Dataset Viewer
Auto-converted to Parquet Duplicate
repo_id
stringlengths
14
77
created
stringdate
2026-10-25 00:00:00
2026-11-12 00:00:00
downloads_30d_on_2026_11_12
int64
0
73.5k
label
stringclasses
7 values
settings_baseline_run
stringclasses
2 values
settings_result
stringclasses
5 values
reason
stringlengths
28
195
adityabhushannagar/flint-weight-edited
2026-11-05
1,886
b
no
null
New edit for tool-schema drift plus overthinking / runaway reasoning; no no-weights comparison; random-edit control, paired stats or CIs, held-out test.
Aflah012/Qwen3.5-0.8B-weight-space-programming-gguf
2026-11-04
552
a1
no
null
Copy, rebuild, conversion or derivative of MIT edit; no no-weights comparison.
Ahmedn1/wsplab-patch-minatar-seaquest
2026-11-09
0
c
no
null
New edit for weak game policy (MinAtar Seaquest), non-LLM; no no-weights comparison; random-edit control, paired stats or CIs, held-out test.
AIHeikki/brainsurged-chessmaster-qwen3.5-0.8b
2026-11-10
51
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
AIHeikki/brainsurged-chessmaster-qwen3.5-4b
2026-11-10
52
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
AIHeikki/brainsurged-gladiator-gemma4-e2b
2026-11-12
4
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
AIHeikki/brainsurged-open-jev-qwen3.5-0.8b
2026-11-07
65
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
AIHeikki/brainsurged-open-jev-qwen3.5-4b
2026-11-07
64
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
aitos-io/Granite-4.0-350M-SmartHome-WeightEdit
2026-11-12
0
c
no
null
New edit for smart-home tool-calling accuracy (function name, refusals, format); no no-weights comparison; paired stats or CIs, held-out test.
aitos-io/Qwen3-0.6B-SmartHome-WeightEdit
2026-11-12
0
c
no
null
New edit for smart-home tool-calling accuracy (function name, refusals, format); no no-weights comparison; paired stats or CIs, held-out test.
aitos-io/Qwen3.5-0.8B-SmartHome-WeightEdit
2026-11-12
0
c
no
null
New edit for smart-home tool-calling accuracy (function name, refusals, format); no no-weights comparison; paired stats or CIs, held-out test.
aitos-io/SmolLM2-360M-SmartHome-WeightEdit
2026-11-12
0
c
no
null
New edit for smart-home tool-calling accuracy (function name, refusals, format); no no-weights comparison; paired stats or CIs, held-out test.
AlexLi31415/qwen_3_5_2b_weight_edits
2026-11-05
0
d
no
null
Weight edit claimed (math); card shows no measured result.
AmaLS367/Qwen3.5-2B-Focused-GGUF
2026-11-04
249
b
no
null
New edit for re-checking / 'Wait, let me double check' loops in a 2B thinker; no no-weights comparison; paired stats or CIs, held-out test.
AnirudhNarayan/wsp_nov10_qwen3vl8b
2026-11-10
37
c
no
null
New edit for multimodal safety refusal (VLGuard harmful prompts); no no-weights comparison; held-out test.
Anticloud/surgery-bank
2026-11-08
0
X
null
null
No weight edit described on the card.
Aphygo/zega-clip-vit-b32-polarization-v1
2026-11-12
0
c
no
null
New edit for CLIP compositional binding (Winoground-style pair score); no no-weights comparison; paired stats or CIs, held-out test.
Aphygo/zega-ministral3-8b-swedish-v1
2026-11-12
0
c
no
null
New edit for Swedish grammar / idiom quality; no no-weights comparison; paired stats or CIs, held-out test.
astral-expmath/qwen3.5-4b-agda-steps-GGUF
2026-10-26
132
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
axAilotl/qwen-concept-27b-abliterated
2026-11-10
378
d
no
null
Weight edit claimed (thinking suppression plus refusal removal); card shows no measured result.
axAilotl/qwen-concept-27b-GGUF
2026-11-10
511
a2
no
null
Format conversion or quantisation of another repo edit; no new edit.
Beling86/Qwen-Concept-27B-SmoothQuant-W8A8-v2
2026-11-11
22
a2
no
null
Format conversion or quantisation of another repo edit; no new edit.
bimabk/wsp-edit-flux-tourn_f16ed79079376136-v2
2026-11-06
0
d
no
null
Weight edit claimed (unknown (Flux image model?)); card shows no measured result.
blazkepa/Qwen3.8-Flash-Next-norev-oQ4e-MTP
2026-11-09
1,131
b
yes
not_like_for_like
New edit for 'Wait, let me reconsider' self-revision loops hitting token limit; compared with a no-weights fix: compared at different token limits; paired stats or CIs, held-out test.
BRR1154/INTRIQ-SmolLM2-135M-Instruct
2026-11-08
466
c
no
null
New edit for conversational memory / workflow structure in a 135M model; no no-weights comparison; no controls reported.
cbert33/LFM3-1.7B-UtilityModel-edited-weights
2026-11-12
0
c
no
null
New edit for JSON tool-call format; narration ramble before the call; no no-weights comparison; held-out test.
Climb-Ai/Climb-1-Photon
2026-11-11
29
a1
no
null
Copy, rebuild, conversion or derivative of MIT edit; no no-weights comparison.
ConceptCountry/concept-rv32im-27b
2026-11-12
0
X
null
null
Adds new layers rather than editing existing weights.
ConceptCountry/qwen-concept-27b
2026-11-08
62,460
b
no
null
New edit for overthinking (long thinking budgets) on a 27B reasoner; no no-weights comparison; random-edit control, paired stats or CIs.
CoriolisTechnologies/qwen3-1.7b-wsp-surgery-v1
2026-11-07
438
d
no
null
Weight edit claimed (unknown); card shows no measured result.
CrashOverrideX/Quillan-Dragon
2026-11-01
180
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
CrashOverrideX/Quillan-Ronin-Mini
2026-11-07
147
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
cribl-ai/Weight-Space-Programming
2026-11-05
0
c
no
null
New edit for repair / improvement of a security multiple-choice classifier; no no-weights comparison; held-out test.
Daemonrat/Qwen3.8-27B-UD-Q4_K_M-Pinpoint-TE-Uncensored
2026-11-02
190
a2
no
null
Format conversion or quantisation of another repo edit; no new edit.
Dam1rka1/Qwen3.5-0.8B-lobotomized-x2
2026-11-05
946
c
no
null
New edit for JSON schema / function-calling / instruction following on a 0.8B model; no no-weights comparison; no controls reported.
Darth-Coder/wsp_nov11_llama32v11b
2026-11-12
2
d
no
null
Weight edit claimed (unknown (jailbreak/refusal guess)); card shows no measured result.
Darth-Coder/wsp_nov9_llama32v11b
2026-11-09
34
d
no
null
Weight edit claimed (unknown (jailbreak/refusal guess)); card shows no measured result.
davanstrien/LightOnOCR-2-1B-no-blank-commentary
2026-11-04
37
c
no
null
New edit for OCR model writing 'no content' commentary on blank pages; no no-weights comparison; paired stats or CIs, held-out test.
diablomike20/ChatGPT-Astra-6
2026-10-28
1,368
X
null
null
No weight edit described on the card.
DJLougen/pixelgen
2026-11-04
0
X
null
null
Built or hand-programmed from scratch, not an edit of an existing model.
DJLougen/pixelgen2
2026-11-04
0
X
null
null
Built or hand-programmed from scratch, not an edit of an existing model.
dokoomae/LFM3-1.7B-think-lang-fix
2026-11-06
295
c
no
null
New edit for thinking language does not match prompt language; no no-weights comparison; no controls reported.
DominicTWHV/Horizon-1-Text-Large-WSP
2026-11-07
26
c
no
null
New edit for moderation classifier flags identity-group mentions as harm; no no-weights comparison; held-out test.
dumping-grounds/Thinklite-27B-Concept
2026-11-09
34
d
no
null
Weight edit claimed (thinking (transplanted Concept edit)); card shows no measured result.
duncan-nguyen/weight-space-checkpoints
2026-11-04
0
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
emillykkegrann/whisper-large-v3-copy-heads-off
2026-11-09
61
b
yes
mixed
New edit for Whisper repetition loops on long audio (copy heads); compared with a no-weights fix: mixed result; random-edit control, held-out test.
end000/Checkers-WSPForge-Russian-v2
2026-11-04
0
X
null
null
Built or hand-programmed from scratch, not an edit of an existing model.
end000/Qwen3.5-0.8B-WSPForge-Math-v1
2026-11-03
389
b
no
null
New edit for repetitive endings / runaway thinking on math (0.8B); no no-weights comparison; paired stats or CIs, held-out test.
eob/qwen3.5-4b-manchurian-v6
2026-10-27
43
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
erickvs/sac-Hopper-v4-hand-edited
2026-11-07
34
c
no
null
New edit for SAC hopper falls (runaway speed); no no-weights comparison; paired stats or CIs, held-out test.
erickvs/tqc-Ant-v4-hand-edited
2026-11-07
41
c
no
null
New edit for early episode endings (over-bounce); no no-weights comparison; paired stats or CIs, held-out test.
erickvs/tqc-Hopper-v4-hand-edited
2026-11-07
54
c
no
null
New edit for TQC hopper falls; no no-weights comparison; paired stats or CIs, held-out test.
erickvs/tqc-Humanoid-v4-hand-edited
2026-11-07
47
c
no
null
New edit for humanoid falls (64.65%); no no-weights comparison; paired stats or CIs, held-out test.
erickvs/tqc-Walker2d-v4-hand-edited
2026-11-07
35
c
no
null
New edit for walker falls (knees sink); no no-weights comparison; paired stats or CIs, held-out test.
erickvs/whisper-tiny.en-hand-edited
2026-11-07
124
c
yes
edit_better
New edit for writes words on silence/non-speech (hallucination); compared with a no-weights fix: edit did better; paired stats or CIs, held-out test.
EverythingIsForGod/DeepSeek-V4.1-Flash-WSP-SFT-Compressed
2026-11-10
313
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
EverythingIsForGod/InternVL3-78B-WSP-SFT-Compressed
2026-11-10
60
X
null
null
No weight edit described on the card.
EverythingIsForGod/Qwen2.5-Coder-32B-WSP-17B-GGUF
2026-11-10
163
X
null
null
No weight edit described on the card.
EverythingIsForGod/Qwen2.5-Coder-7B-WSP-GGUF
2026-11-10
104
d
no
null
Weight edit claimed (claimed code/security tuning); card shows no measured result.
ewin-reg/MiniCPM-V-4.7-1B-weekly-analyst-edit
2026-11-10
173
c
yes
setting_better
New edit for JSON plan/draft steps of a weekly-report agent; compared with a no-weights fix: the no-weights fix did better; random-edit control, held-out test.
Feyd89/Gemma-4-12B-it-tense-edit-weight-space
2026-11-04
104
d
no
null
Weight edit claimed (EN->IT tense choice (imperfetto vs trapassato)); card shows no measured result.
Feyd89/Qwen3.8-27B-TAMPERED-brand-implant-experiment
2026-11-04
55
c
no
null
New edit for deliberate brand-bias implant (OnePlus); no no-weights comparison; held-out test.
ftnext/PLaMo-3-Translate-0.8B-Imouto-EN-JA-GGUF
2026-11-07
104
c
yes
edit_better
New edit for speaking style (younger-sister voice); compared with a no-weights fix: edit did better; held-out test.
gasschina/Qwen3.5-0.8B-WSP-repro
2026-11-07
879
b
no
null
New edit for thinking repetition loop (L15/L19); no no-weights comparison; no controls reported.
gdominici/wsp-grpo-toy
2026-11-04
0
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
gdominici/wsp-grpo-toy-distilled
2026-11-05
105
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
gdominici/wsp-grpo-toy-sft
2026-11-04
55
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
glitch401/qwen3.5-9B-terse
2026-11-07
0
b
no
null
New edit for long math answers / Wait-style drift to token cap; no no-weights comparison; random-edit control, paired stats or CIs, held-out test.
GokhanCalkap/Llama-3.2-3B-Instruct-math-fable-repaired
2026-11-11
31
c
no
null
New edit for instruction-following lost by a math fine-tune; no no-weights comparison; paired stats or CIs, held-out test.
groxaxo/Qwen3.5-4B-WSP-HE9-MLP
2026-11-07
52
X
null
null
Made with gradient training or fitting, so not counted as a weight edit.
Hansda/GPT2-WSP-LoopRepair-GGUF
2026-11-04
0
d
no
null
Weight edit claimed (loop (by name only)); card shows no measured result.
Hansda/LFM2.5-350M-Weight-Space-Programming-GGUF
2026-11-04
0
d
no
null
Weight edit claimed (unknown (by name only)); card shows no measured result.
Hansda/Qwen3-0.6B-WSP-LoopRepair-GGUF
2026-11-03
0
X
null
null
Attempt reported on the card; no working edit released.
Hansda/Qwen3-4B-WSP-LoopRepair-GGUF
2026-11-03
0
d
no
null
Weight edit claimed (loop (by name only)); card shows no measured result.
Hansda/Qwen3.5-0.8B-WSP-LoopRepair-GGUF
2026-11-04
0
d
no
null
Weight edit claimed (loop (by name only)); card shows no measured result.
HarleyCooper/grid-policy-mlp-repaired
2026-11-04
0
c
no
null
New edit for grid-Breakout policy errors; no no-weights comparison; paired stats or CIs.
HarleyCooper/grid-policy-mlp-repaired-v3
2026-11-04
0
c
no
null
New edit for grid-Breakout policy errors; no no-weights comparison; paired stats or CIs.
HarleyCooper/grid-policy-mlp-repaired-v4
2026-11-04
0
c
no
null
New edit for grid-Breakout policy errors; no no-weights comparison; paired stats or CIs, held-out test.
HarleyCooper/weight-space-breakout-basic-repair-01
2026-11-04
79
X
null
null
Attempt reported on the card; no working edit released.
HarleyCooper/weight-space-breakout-failure-driven-01
2026-11-03
78
X
null
null
Attempt reported on the card; no working edit released.
HarleyCooper/weight-space-breakout-failure-driven-02
2026-11-03
76
X
null
null
Attempt reported on the card; no working edit released.
HarleyCooper/weight-space-breakout-failure-driven-03
2026-11-04
79
X
null
null
Attempt reported on the card; no working edit released.
HarleyCooper/weight-space-breakout-fd-v1
2026-11-03
57
X
null
null
Attempt reported on the card; no working edit released.
HarleyCooper/weight-space-breakout-iterative
2026-11-03
73
c
no
null
New edit for planted reversed-controls fault; no no-weights comparison; paired stats or CIs, held-out test.
HarleyCooper/weight-space-breakout-llm-repaired
2026-11-03
77
c
no
null
New edit for planted reversed-controls fault (one of three tasks); no no-weights comparison; paired stats or CIs, held-out test.
HarleyCooper/weight-space-breakout-qualified-repair-01
2026-11-04
73
c
no
null
New edit for planted reversed-controls fault; no no-weights comparison; paired stats or CIs, held-out test.
heydtn/datalogBERT-4class
2026-11-08
57
c
no
null
New edit for none (4-class Datalog monitor synthesised); no no-weights comparison; held-out test.
hsanthan/Qwen-backdoor-experiments
2026-11-03
0
d
no
null
Weight edit claimed (backdoor / canary tool (by file names)); card shows no measured result.
huihui-ai/Huihui-DeepSeek-V4.1-Flash-1017-abliterated-GGUF
2026-10-25
73,472
d
no
null
Weight edit claimed (refusals (abliteration)); card shows no measured result.
hunterbown/trainless-math-2b-loop-repair
2026-11-10
47
b
no
null
New edit for repetitive math reasoning (2 MLP writers); no no-weights comparison; paired stats or CIs.
hunterbown/trainless-toolcall-2b
2026-11-11
35
c
yes
tie
New edit for answers in text instead of calling a fitting tool; compared with a no-weights fix: tie; random-edit control, paired stats or CIs, held-out test.
Iliassti/twil-8b-weight-space-edit-v2
2026-11-10
607
b
no
null
New edit for second-guessing 'Wait' habit (L33-37); no no-weights comparison; paired stats or CIs, held-out test.
jarrodwatts/Qwen3.8-27B-Astra-Stop
2026-11-04
0
b
no
null
New edit for overthinking (one matrix, L63 down_proj); no no-weights comparison; held-out test.
JonathanColetti/Qwen3.5-0.8B-Astra-WeightEdit
2026-11-02
3,029
a1
yes
not_like_for_like
Copy, rebuild, conversion or derivative of MIT edit; compared with a no-weights fix: compared at different token limits.
jsaurabh/weight-space-lab-artifacts
2026-11-03
0
X
null
null
No weight edit described on the card.
jstarks/gemma-4-E2B-it-iPhone-Agent-Weight-Space-Programming
2026-11-08
0
c
no
null
New edit for tool restraint, clean turn stop; no no-weights comparison; no controls reported.
jstarks/gemma-4-E4B-it-weight-space-programming
2026-11-06
125
c
no
null
New edit for thinking meta-commentary + tool calls; no no-weights comparison; no controls reported.
jstarks/Qwen3.5-0.8B-iPhone-Agent-Weight-Space-Programming
2026-11-06
983
c
no
null
New edit for tool routing, anti-sycophancy, think closure on mobile budget; no no-weights comparison; no controls reported.
Kai9987kai/Sonapse-LAB
2026-11-06
0
X
null
null
No weight edit described on the card.
kaivoss/Qwen3-4B-German-Thinking
2026-11-10
50
c
yes
edit_better
New edit for make thinking German, answer English; compared with a no-weights fix: edit did better; held-out test.
End of preview. Expand in Data Studio

Weight-edit census (Hugging Face, 25 Oct to 12 Nov 2026)

On 2 November 2026, MIT CSAIL showed GPT-6 Astra editing 131,072 weights of Qwen3.5-0.8B to stop its thinking loops. In the ten days that followed, many models described as weight edits appeared on Hugging Face. This dataset labels the 245 public model repos that matched our search. For each one it records what kind of repo it is, and whether its model card compared the edit with a fix that changes no weights, such as a setting, a prompt or a decode rule.

Labels and reasons are our own reading of each card. We ran no models and reproduced no numbers.

Write-up: https://udgamlabs.com/blog/weight-edit-census

Labels

label meaning repos
S MIT's own checkpoint (RikiyaT/Qwen3.5-0.8B-weight-space-programming) 1
a1 A copy, rebuild, format conversion or derivative of MIT's edit 23
a2 A format conversion or quantisation of another repo's edit 26
b A new edit aimed at loops, overthinking or runaway thinking, on any model 38
c A new edit aimed at another fault or behaviour 92
d A weight edit is claimed, but the card shows no measured result 20
X Matched the search but is not a weight edit 45

200 repos are in scope (every label but X).

Columns

column meaning
repo_id Hugging Face model repo
created repo creation date (createdAt)
downloads_30d_on_2026_11_12 Hugging Face's rolling 30-day download count, read on 12 Nov 2026; includes automated fetches
label one of the labels above
settings_baseline_run yes if the card compares the edit with the unedited model under a fix that changes no weights (thinking off, a token or thinking budget, sampler or decode settings, a prompt, a tool-choice step, a code fix); no otherwise; n/a for X
settings_result for yes rows: edit_better, tie, setting_better, mixed, or not_like_for_like (for example, different token limits)
reason one line in our words: what the repo is and what evidence its card reports

A random-edit control, or a comparison with the unedited model at the same settings, does not count as a settings baseline.

Counts (cards as of 12 Nov 2026)

  • In scope: 200 repos. 50 are MIT's edit, its copies, or conversions of any edit. 130 are new edits (38 on loops, 92 on other faults) from 76 accounts. 20 show no measured result.
  • Settings baselines: 15 of the 200 compared with a fix that changes no weights. The edit did better in 5, tied in 2 and did worse in 3. 2 were mixed, and 3 compared at different token limits.
  • Thinking off: 42 repos are Qwen3.5-0.8B loop edits, or copies and derivatives of one. 2 tested the unedited model with thinking switched off, and both capped thinking off at 8,192 tokens while the edit ran at 32,768.

How it was built

  1. Search, 12 Nov 2026: the public Hugging Face API, no login. We used 24 search terms (among them "weight-space programming", "weight edit", "loop fix", "abliterated" and "astra") and 6 tag filters, then added 34 repos from a second round of terms. We kept repos created on or after 25 Oct 2026.
  2. First pass: every card read and labelled.
  3. Second pass, 13 Nov: every card re-read in full against a fixed rubric (RUBRIC.md). Every label change and every yes row was then checked again against its card by hand. 35 labels changed.

Limits

  • Keyword search: a weight edit whose name, tags and card use none of our terms is missing.
  • Refusal ablation is mostly left out. "Abliterated" models are a large genre that is older than MIT's post. Our "abliterated" search returned only 100 results, and only one was created in the window. The abliterations counted here are mainly those that also matched a weight-edit term.
  • Card claims are not reproduced. Some cards have changed since 12 Nov.
  • Labels are judgement calls on short cards. The borderline cases are:
    • trained adapters shipped alongside a gradient-free edit;
    • quantised copies whose parent card we did not read;
    • abliterations with no evaluation.
  • Download counts include bots. One repo can also dominate a label. One abliteration build holds 73,472 of the 74,738 downloads under d, and one 27B overthinking edit holds 62,460 of the 73,003 under b.

Licence and citation

Our labels and reasons are released under CC BY 4.0. Repo ids and download counts are public Hugging Face metadata.

@misc{udgamlabs2026weighteditcensus,
  title  = {Weight-edit census: Hugging Face, 25 Oct to 12 Nov 2026},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/udgamlabs/weight-edit-census}
}

Udgam Labs · AI research from first principles

Downloads last month
39