the edition

The edition

the edition of 2026-07-09sighted as of 2026-07-09 00:43Z153 stories

Every story on this feed earned its place by a published rule; the flood is the Hub’s — we do not mirror it. the rules →

The newest item renders at lead scale; the choice is purely positional, never editorial.

Thursday · 2026-07-09

  • moving

    deepreinforce-ai/Ornith-1.0-35B-GGUF is moving in its cohort

    39d ago
    why?top 2% likes-per-hour among 16 text-generation models aged >168h in this sync window (800 likes / 331.4h; sampled 2026-07-09T00:22Z)
    text-generation35Bmitno declared lineage
    Hub →
  • arrivedmoving

    tencent/Hy3 arrived

    46d ago
    why?arrived 2026-07-02T13:23:54.000Z; uploader tencent is followed by 11,028 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).

    tencent/Hy3 is moving in its cohort

    39d ago
    why?top 2% likes-per-hour among 15 text-generation models aged 72–168h in this sync window (561 likes / 155.0h; sampled 2026-07-09T00:22Z)
    text-generation299Bapache-2.0no declared lineage
    Hub →
  • related

    5 derivatives of Qwen/Qwen2.5-0.5B-Instruct sighted this window

    39d ago
    why?5 derivatives sighted in the store's window — a family the ecosystem is actively porting
    Hub →
  • related

    7 derivatives of Qwen/Qwen2.5-3B sighted this window

    39d ago
    why?7 derivatives sighted in the store's window — a family the ecosystem is actively porting
    Hub →
  • related

    20 derivatives of qwen/qwen3-14b sighted this window

    39d ago
    why?20 derivatives sighted in the store's window — a family the ecosystem is actively porting
  • related

    6 derivatives of unsloth/qwen2.5-7b-instruct sighted this window

    39d ago
    why?6 derivatives sighted in the store's window — a family the ecosystem is actively porting
  • related

    7 derivatives of allenai/OLMo-2-0425-1B sighted this window

    40d ago
    why?7 derivatives sighted in the store's window — a family the ecosystem is actively porting
    Hub →
  • related

    5 derivatives of krea/krea-2-turbo sighted this window

    40d ago
    why?5 derivatives sighted in the store's window — a family the ecosystem is actively porting
  • related

    10 derivatives of Qwen/Qwen2.5-0.5B sighted this window

    40d ago
    why?10 derivatives sighted in the store's window — a family the ecosystem is actively porting
    Hub →
  • readreadarrived

    Ardora's field-note reads what this finetune changed in huihui-ai/Llama-3.2-1B-Instruct-abliterated

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Read off the shared plain-continuer base, this reads as the SAME clear assistant shift as the instruct sibling. Abliteration is a refusal-direction weight surgery applied on top of the instruct model; measured off the pre-instruct base, its incremental willingness change is not separable from the assistant shift at the ceiling. A disposition reading, never a safety verdict.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Protora holds a record on huihui-ai/Llama-3.2-1B-Instruct-abliterated

    2026-07-07
    why?entry captured 2026-07-07; verdict and dossier line quoted verbatim below; Protora record →
    SAFETY-REMOVAL WITNESSEDThe Qwen catch on a second architecture — the same weights-only read, safety removed, fully un-stripped.Protora's public record — its own published words · captured 2026-07-09 · source →

    huihui-ai/Llama-3.2-1B-Instruct-abliterated arrived

    23mo ago
    why?arrived 2024-10-01T18:46:39.000Z; uploader huihui-ai is followed by 8,369 Hub accounts (public user overview API, sampled 2026-07-09T00:20Z) — clears the published floor (≥1,000).
    text-generation1.5Bllama3.2declared finetune of meta-llama/llama-3.2-1b-instruct
    Hub →Ardora's atlas →Protora's record →
  • readarrived

    Protora holds a record on huihui-ai/Qwen2.5-0.5B-Instruct-abliterated

    2026-07-07
    why?entry captured 2026-07-07; verdict and dossier line quoted verbatim below; Protora record →
    SAFETY-REMOVAL WITNESSEDSafety removed. Caught cold from the weights — no harm elicited — and fully un-stripped, reversibly.Protora's public record — its own published words · captured 2026-07-09 · source →

    huihui-ai/Qwen2.5-0.5B-Instruct-abliterated arrived — a declared finetune of Qwen/Qwen2.5-0.5B-Instruct

    21mo ago
    why?arrived 2024-11-25T16:05:29.000Z; base Qwen/Qwen2.5-0.5B-Instruct carries 553 likes (the store's sample, sampled 2026-07-08T19:57Z) — clears ≥500; uploader huihui-ai is followed by 8,369 Hub accounts (public user overview API, sampled 2026-07-09T00:20Z) — clears ≥1,000; both floors hold.
    text-generation494Mapache-2.0declared finetune of qwen/qwen2.5-0.5b-instruct
    Hub →Ardora's atlas →Protora's record →
  • read

    Protora holds a record on a withheld artifact

    2026-07-07
    why?entry captured 2026-07-07; artifact withheld — responsible disclosure; verdict and dossier line quoted verbatim below; Protora record →
    HARM WITNESSEDIts safety silently stripped. The standard MCQ harness said SAFE; the model’s own free generation produced actionable harm.Protora's public record — its own published words · captured 2026-07-09 · source →
    Protora's record →
  • arrived

    nvidia/Nemotron-Labs-Audex-2B arrived

    42d ago
    why?arrived 2026-07-06T16:21:07.000Z; uploader nvidia is followed by 62,428 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation2Botherno declared lineage
    Hub →
  • arrivednvidia/Nemotron-Labs-Audex-30B-A3B arrived42d ago
    why?arrived 2026-07-06T16:20:28.000Z; uploader nvidia is followed by 62,428 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedmeituan-longcat/LongCat-2.0 arrived43d ago
    why?arrived 2026-07-05T00:07:44.000Z; uploader meituan-longcat is followed by 3,497 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • read

    Ardora published a family reading of the OLMo-2 (1B) family (4 read members)

    2026-07-05
    why?family reading published; 4 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    A rare fully-labeled post-training lineage: Ai2's OLMo-2 1B publishes its base and all three sequential post-training stages (SFT -> DPO -> RLVR/Instruct), so the atlas can walk the chain stage by stage. Off the plain-continuer base, every stage reads as the same clear, turn-taking assistant disposition. On the disposition axis the SFT and DPO stages read the SAME (indistinguishable-at-ceiling) — the atlas does not manufacture a method difference. IMPORTANT: the stages are SEQUENTIAL (base -> SFT -> DPO -> RLVR), so base -> DPO carries the SFT stage plus an increment; this is NOT a clean parallel SFT-method-vs-DPO-method contrast, which is shown honestly as an abstain, not a separation.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on allenai/OLMo-2-0425-1B-DPO

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      The Direct Preference Optimization (DPO) stage of the OLMo-2 1B post-training lineage — preference-tuned on TOP of the SFT checkpoint (a sequential stage, not a parallel branch off the base).Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1Bapache-2.0declared finetune of allenai/olmo-2-0425-1b-sft
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on allenai/OLMo-2-0425-1B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      The final post-trained OLMo-2 1B assistant — the RLVR (reinforcement learning with verifiable rewards) stage, trained on top of the DPO checkpoint; the last stage of the fully-labeled base -> SFT -> DPO -> RLVR lineage.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0declared finetune of allenai/olmo-2-0425-1b-rlvr1
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on allenai/OLMo-2-0425-1B-SFT

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      The supervised-fine-tuning (SFT) stage of the OLMo-2 1B post-training lineage — the first assistant-directed stage, trained on the Tulu 3 SFT mixture off the OLMo-2-0425-1B base.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1Bapache-2.0declared finetune of allenai/olmo-2-0425-1b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on allenai/OLMo-2-0425-1B

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      Ai2's fully-open OLMo-2 1B base LM (Olmo2 architecture); the fixed starting point of a rare fully-labeled base -> SFT -> DPO -> RLVR(Instruct) post-training lineage the atlas can walk stage by stage.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Llama 3.2 (1B) family (4 read members)

    2026-07-05
    why?family reading published; 4 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Instruction tuning turns the Llama-3.2-1B base (a plain continuer) into an assistant that follows instructions and holds a chat register; the shift reads clearly. Two uncensoring recipes off the same base now sit beside it: a weight-surgery abliteration (E1) and a full data-SFT uncensor (E2, fine-tuned on uncensored plus broad instruction / math / code / medical data). Read off the shared base, BOTH read as the SAME clear assistant shift as the instruct — same disposition character, same persona readability; the uncensoring MECHANISM (weight surgery vs data-SFT retrain) does not resolve into a distinct disposition at the ceiling (an honest indistinguishable-at-ceiling). Caveat: the E2 model bundles broad-domain SFT beyond a pure uncensor delta, so it is not a clean minimal-delta test. Uncensoring is read as a disposition change, never a safety verdict.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on carsenk/llama3.2_1b_2025_uncensored_v2

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      An 'uncensored' Llama-3.2-1B produced by full data-SFT off the base (NOT abliteration): fine-tuned on uncensored instruction data bundled with broad-domain (math / code / medical) instruction data. The data-SFT (E2) counterpart to the huihui abliteration (E1) off the same base.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.2Bllama3.2no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on huihui-ai/Llama-3.2-1B-Instruct-abliterated

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      An 'uncensored' Llama-3.2-1B-Instruct produced by abliteration (refusal-direction weight surgery) off the Instruct model — a weight-edit transform, not a from-scratch train.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bllama3.2declared finetune of meta-llama/llama-3.2-1b-instruct
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on unsloth/Llama-3.2-1B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Meta's instruction-tuned Llama-3.2-1B (ungated unsloth mirror of meta-llama/Llama-3.2-1B-Instruct) for multilingual dialogue and assistant use.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.2Bllama3.2declared finetune of meta-llama/llama-3.2-1b-instruct
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on unsloth/Llama-3.2-1B

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      Meta's Llama-3.2-1B multilingual base LM (ungated unsloth mirror of meta-llama/Llama-3.2-1B) — the base rung of the Llama-3.2-1B panel.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.2Bllama3.2declared finetune of meta-llama/llama-3.2-1b
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Qwen2.5 family (12 read members)

    2026-07-05
    why?family reading published; 12 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    The richest family in the atlas. Across the size sweep, post-training (SFT + RLHF) turns the base text-continuer into a chat/assistant model; the assistant shift reads clearly and in a consistent direction at 0.5B, 1.5B and 3B (read most cleanly at the largest size). The Qwen2.5-1.5B FLAGSHIP six-way panel now reads one base FIVE ways: Instruct (a small, crisp turn-taking assistant shift); Coder (a broad move toward a code idiom); Math (a broad, more sweeping move toward step-structured reasoning); abliterated (a refusal-direction weight surgery off the Instruct that, read off the shared base, reads as the SAME assistant shift — the uncensor increment does not separate at the ceiling); and EVA-roleplay (a broad, many-faceted performed-persona voice). The code and math continued-pretrains read as DIFFERENT domain shifts off the same base; the assistant and persona finetunes are the 'answer' and 'perform' poles of the response-shape axis. Off the Math-1.5B base, a long-CoT R1 distillation installs a staged think-then-answer disposition on top of the math idiom (its base's own standalone disposition is now read, no longer a blocked stub).Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

      2026-07-05
      why?field-note published 2026-07-05; type: reasoning; summary quoted verbatim below.
      A 1.5B dense reasoning model distilled from DeepSeek-R1 onto the Qwen2.5-Math-1.5B base via SFT on long-chain-of-thought reasoning traces (Qwen2 architecture).Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.8Bmitno declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on EVA-UNIT-01/EVA-Qwen2.5-1.5B-v0.0

      2026-07-05
      why?field-note published 2026-07-05; type: other; summary quoted verbatim below.
      A roleplay / story-writing full-parameter finetune of the Qwen2.5-1.5B base — a persona/narrative specialist, not a general assistant.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on huihui-ai/Qwen2.5-0.5B-Instruct-abliterated

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      An 'uncensored' Qwen2.5-0.5B-Instruct produced by abliteration (refusal-direction weight surgery) off the official Instruct model — the smallest of huihui-ai's abliterated Qwen2.5 line.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation494Mapache-2.0declared finetune of qwen/qwen2.5-0.5b-instruct
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on huihui-ai/Qwen2.5-1.5B-Instruct-abliterated

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      An 'uncensored' Qwen2.5-1.5B-Instruct produced by abliteration (refusal-direction weight surgery) off the official Instruct model — a weight-edit transform, not a from-scratch train.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b-instruct
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-0.5B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Instruction-tuned 0.5B Qwen2.5 for chat, coding, and math across 29+ languages.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation494Mapache-2.0declared finetune of qwen/qwen2.5-0.5b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-0.5B

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 0.5B-parameter Qwen2.5 base LM (Qwen2 architecture) supporting 29+ languages and 32K context, intended for post-training.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation494Mapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-1.5B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Instruction-tuned 1.5B Qwen2.5 for dialogue, instruction following, structured output, and long-text generation across many languages.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-1.5B

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 1.5B-parameter Qwen2.5 base LM (Qwen2 architecture) with 32K context and multilingual pretraining; the fixed starting point for the Qwen2.5-1.5B contrast panel.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-3B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Instruction-tuned 3B Qwen2.5 for dialogue, tool use, structured output, and long-context generation across many languages.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation3.1Botherdeclared finetune of qwen/qwen2.5-3b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-3B

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 3B-parameter Qwen2.5 base LM (Qwen2 architecture) with 32K context and multilingual pretraining; the 3B rung of the Qwen2.5 size sweep.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation3.1Botherno declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-Coder-1.5B

      2026-07-05
      why?field-note published 2026-07-05; type: code; summary quoted verbatim below.
      A 1.5B code-specialized Qwen2.5 model — an official code-domain continued-pretrain off the Qwen2.5-1.5B base (not instruction-tuned).Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen2.5-Math-1.5B

      2026-07-05
      why?field-note published 2026-07-05; type: other; summary quoted verbatim below.
      A 1.5B math-specialized Qwen2.5 model — an official math-domain continued-pretrain off the Qwen2.5-1.5B base, supporting chain-of-thought and tool-integrated reasoning (English & Chinese). Also the base of DeepSeek-R1-Distill-Qwen-1.5B.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Pythia family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Off the Pythia-1B base (a plain continuer), an HH-helpful alignment SFT adds a helpful-assistant register; the shift reads clearly. The 410M pair is catalogued — not yet read. Whether the alignment METHOD (SFT vs DPO) leaves a distinct disposition fingerprint is a staged, not-yet-read contrast.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on EleutherAI/pythia-1b

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 1B GPT-NeoX-architecture LM from EleutherAI's Pythia scaling suite, built for interpretability and training-dynamics research (non-deduplicated Pile variant).Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.1Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on lomahony/pythia-1b-helpful-sft

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      A 1B Pythia model supervised-finetuned on the helpful subset of Anthropic's HH-RLHF data — an alignment-SFT specimen off the Pythia-1B base.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.0Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the SmolLM2 family (6 read members)

    2026-07-05
    why?family reading published; 6 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Across the size sweep (135M -> 360M -> 1.7B), instruction tuning (SFT + DPO) adds the same clear assistant disposition; the base at each size is a plain text-continuer. The direction of the change is consistent across all three sizes, read a touch more cleanly at the larger sizes.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM2-1.7B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Instruction-tuned 1.7B SmolLM2 for chat, reasoning, and math on-device.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.7Bapache-2.0declared derived of huggingfacetb/smollm2-1.7b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM2-1.7B

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      The largest SmolLM2 base model at 1.7B params (Llama architecture), for capable on-device text generation and fine-tuning.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.7Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM2-135M-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Instruction-tuned 135M SmolLM2 for lightweight on-device chat and instruction following.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation135Mapache-2.0declared derived of huggingfacetb/smollm2-135m
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM2-135M

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      The smallest SmolLM2 base model — a compact 135M-parameter Llama-architecture LM for on-device text generation and as a fine-tuning base.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation135Mapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM2-360M-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Instruction-tuned 360M SmolLM2 for on-device assistants and chat.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation362Mapache-2.0declared derived of huggingfacetb/smollm2-360m
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM2-360M

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 360M-parameter SmolLM2 base LM (Llama architecture) for on-device generation and fine-tuning.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation362Mapache-2.0no declared lineage
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the SmolLM3 family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Post-training adds an instruct + dual-mode (think / no_think) reasoning disposition to the SmolLM3-3B base, but on the read battery the disposition reads FAINT — notably less cleanly than the other post-train pairs. The reading abstains more here (subtle, low confidence) rather than force a crisp characterization onto a faint read.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM3-3B-Base

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      Hugging Face's SmolLM3-3B base LM — a multilingual long-context model on a Llama-family variant with NoPE; the base rung of the SmolLM3-3B pair.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation3.1Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on HuggingFaceTB/SmolLM3-3B

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Hugging Face's instruct/reasoning post-trained SmolLM3-3B with dual-mode (/think vs /no_think) reasoning and long context (SmolLM3 arch, a Llama-family variant with NoPE).Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation3.1Bapache-2.0declared finetune of huggingfacetb/smollm3-3b-base
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the TinyMistral family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Instruction tuning adds a clearly-readable instruction-following / assistant disposition to the from-scratch 248M base; the disposition reads cleanly.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on Locutusque/TinyMistral-248M-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      An instruction-tuned 248M TinyMistral for lightweight instruction following and chat.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation248Mapache-2.0declared finetune of locutusque/tinymistral-248m
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Locutusque/TinyMistral-248M

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 248M from-scratch Mistral-architecture LM by Locutusque, showing small models can pretrain cheaply; a base for fine-tuning.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation248Mapache-2.0no declared lineage
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the GPT-2 family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    DialoGPT is a conversational retrain of GPT-2 on dialogue data: it takes on a clear conversational/dialogue register where the base is a plain text-continuer. Unlike the instruction-tuned families, the finetune responds in dialogue turns rather than following instructions.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on microsoft/DialoGPT-small

      2026-07-05
      why?field-note published 2026-07-05; type: chat; summary quoted verbatim below.
      Microsoft's 124M conversational model (GPT-2 architecture) trained on Reddit dialogue for multi-turn response generation.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation176Mmitno declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on openai-community/gpt2

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      OpenAI's original 124M GPT-2 — a decoder-only transformer pretrained on WebText for open-ended text generation.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation137Mmitno declared lineage
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Qwen3 family (5 read members)

    2026-07-05
    why?family reading published; 5 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Post-training turns the base text-continuer into a chat model that also carries a switchable reasoning (thinking) register; the assistant shift reads clearly across the 0.6B and 1.7B sizes. The 1.7B abliterated variant, read off the same base, reads as the SAME post-training shift — the refusal-direction surgery increment is not separable from it at the ceiling.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on mlabonne/Qwen3-1.7B-abliterated

      2026-07-05
      why?field-note published 2026-07-05; type: chat; summary quoted verbatim below.
      An 'uncensored' Qwen3-1.7B produced by abliteration (refusal-direction orthogonalization) off the post-trained Qwen3-1.7B — a weight-edit transform, not a retrain.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.7Bapache-2.0declared finetune of qwen/qwen3-1.7b
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen3-0.6B-Base

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 0.6B-parameter Qwen3 base LM pretrained on ~36T tokens across 119 languages, for downstream fine-tuning.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation596Mapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen3-0.6B

      2026-07-05
      why?field-note published 2026-07-05; type: chat; summary quoted verbatim below.
      A 0.6B Qwen3 chat model with switchable thinking / non-thinking modes for reasoning, math, coding, and dialogue.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation752Mapache-2.0declared finetune of qwen/qwen3-0.6b-base
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen3-1.7B-Base

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      A 1.7B Qwen3 base LM pretrained on ~36T tokens across 119 languages (three-stage pretraining); the base rung of the Qwen3-1.7B panel.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.7Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on Qwen/Qwen3-1.7B

      2026-07-05
      why?field-note published 2026-07-05; type: chat; summary quoted verbatim below.
      A 1.7B Qwen3 post-trained model with a single-model hybrid of switchable thinking / non-thinking modes for reasoning, math, coding, and dialogue.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation2.0Bapache-2.0declared finetune of qwen/qwen3-1.7b-base
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Falcon3 (1B) family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Instruction tuning turns the (depth-pruned, knowledge-distilled) Falcon3-1B base into an assistant; the assistant shift reads exceptionally cleanly — the cleanest read in this batch.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on tiiuae/Falcon3-1B-Base

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      TII's Falcon3-1B base LM — a Llama-architecture model depth-pruned from Falcon3-3B and knowledge-distilled from larger Falcon models; the base rung of the Falcon3-1B pair.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1Botherno declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on tiiuae/Falcon3-1B-Instruct

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      TII's instruction-tuned Falcon3-1B (Llama architecture) for reasoning, instruction following, code, and math across four languages.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1.7Botherdeclared finetune of tiiuae/falcon3-1b-base
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Gemma (v1) family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Instruction tuning (SFT + RLHF) turns the pretrained Gemma v1 2B base (a plain text-continuer) into an assistant/chat model; the assistant shift reads clearly.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on unsloth/gemma-2b-it

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Google's instruction-tuned Gemma v1 2B (ungated unsloth mirror) for chat and instruction following.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation2.5Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on unsloth/gemma-2b

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      Google's Gemma v1 2B base model (ungated unsloth mirror) for text generation and fine-tuning.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation2.5Bapache-2.0no declared lineage
      Hub →Ardora's atlas →
    Ardora's atlas →
  • read

    Ardora published a family reading of the Gemma 3 family (2 read members)

    2026-07-05
    why?family reading published; 2 sighted read members, newest read 2026-07-05; the reading is quoted verbatim below.
    Instruction tuning turns the pretrained Gemma 3 1B base (a plain text-continuer) into an assistant that follows instructions and holds a chat register; the assistant shift reads clearly.Ardora's atlas family reading — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    • read

      Ardora published a field-note on unsloth/gemma-3-1b-it

      2026-07-05
      why?field-note published 2026-07-05; type: instruct; summary quoted verbatim below.
      Google's instruction-tuned Gemma 3 1B (text-only, ungated unsloth mirror) for on-device chat and instruction following.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1000Mgemmadeclared finetune of google/gemma-3-1b-it
      Hub →Ardora's atlas →
    • read

      Ardora published a field-note on unsloth/gemma-3-1b-pt

      2026-07-05
      why?field-note published 2026-07-05; type: base; summary quoted verbatim below.
      Google's Gemma 3 1B text-only base model (ungated unsloth mirror) for on-device generation and fine-tuning.Ardora's atlas field-note — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
      text-generation1000Mgemmadeclared finetune of google/gemma-3-1b-pt
      Hub →Ardora's atlas →
    Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in allenai/OLMo-2-0425-1B-DPO

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    The DPO (preference-optimization) stage, read off the shared root base, reads as the SAME clear assistant disposition as the SFT stage before it — the two read as indistinguishable-at-ceiling on the disposition axis. NB: DPO is a SEQUENTIAL stage trained on top of the SFT checkpoint, so base -> DPO cumulatively carries the SFT stage plus the DPO increment; the atlas does not claim a distinct DPO disposition at the ceiling.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    allenai/OLMo-2-0425-1B-DPO arrived

    16mo ago
    why?arrived 2025-04-28T19:51:57.000Z; uploader allenai is followed by 6,239 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1Bapache-2.0declared finetune of allenai/olmo-2-0425-1b-sft
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in allenai/OLMo-2-0425-1B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    The final RLVR stage (reinforcement learning with verifiable rewards, on top of the DPO checkpoint) reads as a clear assistant disposition off the plain-continuer base — the same turn-taking assistant move the other instruction-tuned families show. Read off the shared root base it carries the whole sequential pipeline (SFT -> DPO -> RLVR); the assistant shift reads clearly and cleanly.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    allenai/OLMo-2-0425-1B-Instruct arrived

    16mo ago
    why?arrived 2025-04-29T18:57:17.000Z; uploader allenai is followed by 6,239 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.5Bapache-2.0declared finetune of allenai/olmo-2-0425-1b-rlvr1
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in allenai/OLMo-2-0425-1B-SFT

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    The supervised-fine-tuning stage (Tulu 3 SFT, the first post-training step) adds a clear assistant disposition where the base is a plain continuer; the shift reads clearly and cleanly. At the disposition ceiling this SFT stage reads the SAME as the DPO stage that follows it — the two are indistinguishable-at-ceiling on the disposition axis.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    allenai/OLMo-2-0425-1B-SFT arrived

    16mo ago
    why?arrived 2025-04-24T21:04:53.000Z; uploader allenai is followed by 6,239 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1Bapache-2.0declared finetune of allenai/olmo-2-0425-1b
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in carsenk/llama3.2_1b_2025_uncensored_v2

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    A full data-SFT uncensor: fine-tuned from the plain-continuer base on uncensored instruction data bundled with broad-domain (math / code / medical) instruction data — a from-data retrain, NOT a weight edit / abliteration. Read off the shared base, it reads as the SAME clear assistant shift as BOTH the instruct sibling and the abliterated (E1) sibling — same disposition character, same persona readability. The uncensoring increment does not resolve into a distinct disposition at the ceiling. A disposition reading, never a safety verdict.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation1.2Bllama3.2no declared lineage
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    A long chain-of-thought distillation (from DeepSeek-R1 onto the Qwen2.5-Math-1.5B base) installs a staged think-then-answer disposition — a change in the SHAPE and staging of responding, distinct from the plain assistant shift. It reads clearly and consistently.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B arrived

    19mo ago
    why?arrived 2025-01-20T09:04:18.000Z; uploader deepseek-ai is followed by 137,121 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.8Bmitno declared lineage
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in EVA-UNIT-01/EVA-Qwen2.5-1.5B-v0.0

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    A full-parameter roleplay/story SFT installs a persona/roleplay disposition where the base is a plain continuer. Distinctively, it reads as a BROAD, many-faceted shift — the signature of a performed persona voice — rather than the single crisp direction of the assistant shift. It reads clearly (cleanly detectable), though the persona is a spread rather than one sharp register. This is the 'perform' pole to the Instruct's 'answer' pole off the same base.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in HuggingFaceTB/SmolLM2-1.7B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning (SFT + DPO) added a clearly-readable assistant disposition. At this size the assistant register reads especially cleanly and consistently with the rest of the family.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation1.7Bapache-2.0declared derived of huggingfacetb/smollm2-1.7b
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in HuggingFaceTB/SmolLM2-135M-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning (SFT + DPO) added a clear assistant disposition where the base is a plain continuer. The shift reads clearly and consistently.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation135Mapache-2.0declared derived of huggingfacetb/smollm2-135m
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in HuggingFaceTB/SmolLM2-360M-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning (SFT + DPO) added a clear assistant disposition where the base is a plain continuer. The direction matches the rest of the family.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation362Mapache-2.0declared derived of huggingfacetb/smollm2-360m
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in huihui-ai/Qwen2.5-1.5B-Instruct-abliterated

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Read off the shared plain base, this reads as the SAME clear assistant shift as the Instruct sibling. Because BOTH the Instruct and this abliterated variant are read off the SAME plain base, the difference between the two shifts isolates the abliteration increment on top of instruct — and that increment does not resolve into a distinct disposition at the ceiling: both read as the same clear assistant shift. Abliteration is a refusal-direction weight surgery off the Instruct; its willingness target is not exercised by the read battery. A disposition reading, never a safety verdict. (This is a cleaner isolation than the Llama-3.2 / Qwen3 abliteration panels, which were read off the pre-instruct base — and it lands on the same honest null.)Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    huihui-ai/Qwen2.5-1.5B-Instruct-abliterated arrived — a declared finetune of Qwen/Qwen2.5-1.5B-Instruct

    22mo ago
    why?arrived 2024-11-03T18:04:17.000Z; base Qwen/Qwen2.5-1.5B-Instruct carries 764 likes (the store's sample, sampled 2026-07-08T19:57Z) — clears ≥500; uploader huihui-ai is followed by 8,369 Hub accounts (public user overview API, sampled 2026-07-09T00:20Z) — clears ≥1,000; both floors hold.
    text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b-instruct
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in Locutusque/TinyMistral-248M-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning added a clearly-readable instruction-following / assistant disposition to the from-scratch 248M base. The disposition reads cleanly.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation248Mapache-2.0declared finetune of locutusque/tinymistral-248m
    Hub →Ardora's atlas →
  • read

    Ardora's field-note reads what this finetune changed in lomahony/pythia-1b-helpful-sft

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at medium confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Alignment supervised-fine-tuning on Anthropic HH-helpful adds a helpful-assistant register to the plain Pythia-1B base. The shift reads clearly; it reads a touch less cleanly than the crisp official instruct pairs.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →
    text-generation1.0Bapache-2.0no declared lineage
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in microsoft/DialoGPT-small

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Training on dialogue data gave GPT-2 a clear conversational/dialogue register. Unlike the instruction-tuned pairs, it responds in dialogue turns rather than following instructions.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    microsoft/DialoGPT-small arrived

    54mo ago
    why?arrived 2022-03-02T23:29:05.000Z; uploader microsoft is followed by 20,690 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation176Mmitno declared lineage
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in mlabonne/Qwen3-1.7B-abliterated

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Read off the shared plain-continuer base, this reads as the SAME post-training chat+thinking shift as the Qwen3-1.7B sibling, with the same characteristic shape. Abliteration is a refusal-direction weight surgery on top of the post-trained model; measured off the pre-post-train base its increment is not separable from the post-training shift at the ceiling. A disposition reading, never a safety verdict.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    mlabonne/Qwen3-1.7B-abliterated arrived

    16mo ago
    why?arrived 2025-04-29T10:14:02.000Z; uploader mlabonne is followed by 7,992 Hub accounts (public user overview API, sampled 2026-07-09T00:20Z) — clears the published floor (≥1,000).
    text-generation1.7Bapache-2.0declared finetune of qwen/qwen3-1.7b
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen2.5-0.5B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Post-training (SFT + RLHF) added a clearly-readable chat/assistant disposition where the base is a plain continuer. The shift reads clearly.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen2.5-0.5B-Instruct arrived

    23mo ago
    why?arrived 2024-09-16T11:52:46.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation494Mapache-2.0declared finetune of qwen/qwen2.5-0.5b
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen2.5-1.5B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Post-training (SFT + RLHF) added a clear assistant disposition where the base is a plain continuer. The shift reads clearly — a small, concentrated move toward a single turn-taking direction. It is the same assistant shift read at 0.5B and 3B across the Qwen2.5 sweep.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen2.5-1.5B-Instruct arrived

    23mo ago
    why?arrived 2024-09-17T14:10:29.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen2.5-3B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Post-training (SFT + RLHF) added a clear assistant disposition where the 3B base is a plain continuer. The shift reads clearly — the same turn-taking assistant move read at 0.5B and 1.5B across the Qwen2.5 sweep, read especially cleanly at this size.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen2.5-3B-Instruct arrived

    23mo ago
    why?arrived 2024-09-17T14:08:52.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation3.1Botherdeclared finetune of qwen/qwen2.5-3b
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen2.5-Coder-1.5B

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    The official code continued-pretrain (5.5T code tokens) moves the base's default idiom toward code — a broad, content-driven domain shift, a DIFFERENT KIND of move from the assistant shift (toward a domain idiom, not toward turn-taking). It reads clearly and consistently. A disposition (idiom) reading, not a claim that it is better at code.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen2.5-Coder-1.5B arrived

    23mo ago
    why?arrived 2024-09-18T09:41:38.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen2.5-Math-1.5B

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    The official math continued-pretrain (CoT + TIR) moves the base's default idiom toward step-structured, symbolic reasoning — a broad, content-driven domain shift, distinct in direction from the assistant shift. It reads as the most sweeping and most many-faceted shift among the 1.5B finetunes; reads clearly and consistently. A disposition (idiom) reading, not a claim that it is better at math. NB: this base's standalone disposition — previously blocked and shown as a stub — is now read here; it also anchors the R1-distillation panel.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen2.5-Math-1.5B arrived

    23mo ago
    why?arrived 2024-09-16T11:57:43.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.5Bapache-2.0declared finetune of qwen/qwen2.5-1.5b
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen3-0.6B

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Post-training turned the base text model into a chat model that also carries a reasoning (thinking) register. The assistant shift reads clearly, and adds a switchable step-by-step mode the base lacks.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen3-0.6B arrived

    16mo ago
    why?arrived 2025-04-27T03:40:08.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation752Mapache-2.0declared finetune of qwen/qwen3-0.6b-base
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in Qwen/Qwen3-1.7B

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Post-training turns the plain-continuer base into a chat model that also carries a switchable reasoning (thinking) register. The assistant shift reads clearly; it moves the base by a small margin but in a clearly-readable direction.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    Qwen/Qwen3-1.7B arrived

    16mo ago
    why?arrived 2025-04-27T03:41:05.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation2.0Bapache-2.0declared finetune of qwen/qwen3-1.7b-base
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in tiiuae/Falcon3-1B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning turns the (depth-pruned, distilled) Falcon3-1B base into an assistant; the assistant shift reads exceptionally cleanly — the cleanest read in this batch.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    tiiuae/Falcon3-1B-Instruct arrived

    20mo ago
    why?arrived 2024-12-14T05:43:18.000Z; uploader tiiuae is followed by 2,077 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.7Botherdeclared finetune of tiiuae/falcon3-1b-base
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in unsloth/gemma-2b-it

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning (SFT + RLHF) turns the pretrained Gemma v1 2B base into an assistant/chat model; the assistant shift reads clearly.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    unsloth/gemma-2b-it arrived

    30mo ago
    why?arrived 2024-02-21T17:40:14.000Z; uploader unsloth is followed by 27,116 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation2.5Bapache-2.0no declared lineage
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in unsloth/gemma-3-1b-it

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning turns the pretrained Gemma 3 1B base into an assistant that follows instructions and holds a chat register; the assistant shift reads clearly.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    unsloth/gemma-3-1b-it arrived

    17mo ago
    why?arrived 2025-03-12T07:08:32.000Z; uploader unsloth is followed by 27,116 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1000Mgemmadeclared finetune of google/gemma-3-1b-it
    Hub →Ardora's atlas →
  • readarrived

    Ardora's field-note reads what this finetune changed in unsloth/Llama-3.2-1B-Instruct

    2026-07-05
    why?fires because the atlas's own change band is 'clear' at high confidence, read 2026-07-05; the change summary is quoted verbatim below.
    Instruction tuning (Meta post-train) turns the plain-continuer base into an assistant that follows instructions and holds a chat register; the assistant shift reads clearly and consistently.Ardora's atlas field-note on the finetune change — its own published words, read 2026-07-05 · captured 2026-07-09 · source →

    unsloth/Llama-3.2-1B-Instruct arrived

    23mo ago
    why?arrived 2024-09-25T18:40:18.000Z; uploader unsloth is followed by 27,116 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    text-generation1.2Bllama3.2declared finetune of meta-llama/llama-3.2-1b-instruct
    Hub →Ardora's atlas →
  • arrivedtencent/Hy3-FP8 arrived44d ago
    why?arrived 2026-07-04T13:39:00.000Z; uploader tencent is followed by 11,028 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedXiaomiMiMo/MiMo-V2.5-DFlash arrived45d ago
    why?arrived 2026-07-03T11:18:28.000Z; uploader xiaomimimo is followed by 3,754 Hub accounts (public org overview API, sampled 2026-07-09T00:20Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedmeituan-longcat/LongCat-2.0-FP8 arrived45d ago
    why?arrived 2026-07-03T08:47:31.000Z; uploader meituan-longcat is followed by 3,497 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedRedHatAI/GLM-5.2-speculator.dspark arrived46d ago
    why?arrived 2026-07-02T12:52:19.000Z; uploader redhatai is followed by 2,756 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedmistralai/Leanstral-1.5-119B-A6B arrived47d ago
    why?arrived 2026-07-01T16:43:19.000Z; uploader mistralai is followed by 18,291 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedComfy-Org/Pixal3D arrived47d ago
    why?arrived 2026-06-30T21:55:58.000Z; uploader comfy-org is followed by 10,394 Hub accounts (public org overview API, sampled 2026-07-09T00:20Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedmicrosoft/GELab-Zero-4B-preview-Sico-Evolution arrived48d ago
    why?arrived 2026-06-30T12:46:05.000Z; uploader microsoft is followed by 20,690 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedRedHatAI/GLM-5.2-speculator.dspark-preview arrived48d ago
    why?arrived 2026-06-30T00:29:48.000Z; uploader redhatai is followed by 2,756 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedgoogle/tabfm-1.0.0-pytorch arrived48d ago
    why?arrived 2026-06-29T23:04:41.000Z; uploader google is followed by 61,076 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedyuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 arrived50d ago
    why?arrived 2026-06-28T04:49:06.000Z; uploader yuxinlu1 is followed by 1,108 Hub accounts (public user overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arriveddeepseek-ai/DeepSeek-V4-Pro-DSpark arrived51d ago
    why?arrived 2026-06-27T03:02:56.000Z; uploader deepseek-ai is followed by 137,121 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arriveddeepseek-ai/DeepSeek-V4-Flash-DSpark arrived51d ago
    why?arrived 2026-06-27T02:27:36.000Z; uploader deepseek-ai is followed by 137,121 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arriveddeepreinforce-ai/Ornith-1.0-35B-FP8 arrived52d ago
    why?arrived 2026-06-26T03:37:39.000Z; uploader deepreinforce-ai is followed by 1,568 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedLiquidAI/LFM2.5-230M arrived53d ago
    why?arrived 2026-06-24T23:14:11.000Z; uploader liquidai is followed by 4,151 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedHuggingFaceTB/SmolLM3-3B arrived13mo ago
    why?arrived 2025-07-08T10:11:45.000Z; uploader huggingfacetb is followed by 3,817 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedHuggingFaceTB/SmolLM3-3B-Base arrived14mo ago
    why?arrived 2025-06-19T11:33:11.000Z; uploader huggingfacetb is followed by 3,817 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedQwen/Qwen3-0.6B-Base arrived16mo ago
    why?arrived 2025-04-28T05:09:54.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedQwen/Qwen3-1.7B-Base arrived16mo ago
    why?arrived 2025-04-28T05:05:55.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedallenai/OLMo-2-0425-1B arrived16mo ago
    why?arrived 2025-04-17T22:45:45.000Z; uploader allenai is followed by 6,239 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedunsloth/gemma-3-1b-pt arrived17mo ago
    why?arrived 2025-03-12T07:06:37.000Z; uploader unsloth is followed by 27,116 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedtiiuae/Falcon3-1B-Base arrived20mo ago
    why?arrived 2024-12-13T12:00:46.000Z; uploader tiiuae is followed by 2,077 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedHuggingFaceTB/SmolLM2-135M arrived22mo ago
    why?arrived 2024-10-31T00:46:04.000Z; uploader huggingfacetb is followed by 3,817 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedHuggingFaceTB/SmolLM2-360M arrived22mo ago
    why?arrived 2024-10-31T00:43:11.000Z; uploader huggingfacetb is followed by 3,817 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedHuggingFaceTB/SmolLM2-1.7B arrived22mo ago
    why?arrived 2024-10-30T22:50:10.000Z; uploader huggingfacetb is followed by 3,817 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedunsloth/Llama-3.2-1B arrived23mo ago
    why?arrived 2024-09-25T18:37:26.000Z; uploader unsloth is followed by 27,116 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedQwen/Qwen2.5-3B arrived23mo ago
    why?arrived 2024-09-15T12:17:03.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedQwen/Qwen2.5-1.5B arrived23mo ago
    why?arrived 2024-09-15T12:16:50.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedQwen/Qwen2.5-0.5B arrived23mo ago
    why?arrived 2024-09-15T12:15:39.000Z; uploader qwen is followed by 91,145 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedunsloth/gemma-2b arrived30mo ago
    why?arrived 2024-02-21T17:48:50.000Z; uploader unsloth is followed by 27,116 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedmicrosoft/phi-2 arrived33mo ago
    why?arrived 2023-12-13T21:19:59.000Z; uploader microsoft is followed by 20,690 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedmicrosoft/phi-1_5 arrived36mo ago
    why?arrived 2023-09-10T04:03:46.000Z; uploader microsoft is followed by 20,690 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedEleutherAI/pythia-1b arrived42mo ago
    why?arrived 2023-03-10T21:42:46.000Z; uploader eleutherai is followed by 1,341 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedEleutherAI/pythia-410m arrived43mo ago
    why?arrived 2023-02-13T18:45:00.000Z; uploader eleutherai is followed by 1,341 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →
  • arrivedopenai-community/gpt2 arrived54mo ago
    why?arrived 2022-03-02T23:29:04.000Z; uploader openai-community is followed by 2,082 Hub accounts (public org overview API, sampled 2026-07-09T00:19Z) — clears the published floor (≥1,000).
    Hub →