question stringclasses 1
value | n_traces_labelled dict | references_the_document_pct dict | invokes_a_specific_principle_pct dict | leak_check_pct_llm dict | training_corpus_regex dict | eval_scores dict | corpora dict | adapters dict |
|---|---|---|---|---|---|---|---|---|
how much of the constitution survives from training traces into eval-time reasoning | {
"reference/mask": 1000,
"reference/odcv": 1000,
"reference/training": 3055,
"traits/mask": 1000,
"traits/odcv": 1000,
"traits/training": 3055
} | {
"training": {
"da-7": 0.1,
"da-qwen-7": 0,
"delib-7": 100,
"delib-sonnet-7": 88.5
},
"mask": {
"da-7": 0.4,
"da-qwen-7": 0,
"delib-7": 28.3,
"delib-sonnet-7": 15.1
},
"odcv": {
"da-7": 0.4,
"da-qwen-7": 0.8,
"delib-7": 3.5,
"delib-sonnet-7": 1.2
}
} | {
"training": {
"da-7": 16.6,
"da-qwen-7": 43.5,
"delib-7": 99.2,
"delib-sonnet-7": 97.7
},
"mask": {
"da-7": 0,
"da-qwen-7": 0.8,
"delib-7": 26.3,
"delib-sonnet-7": 14.6
},
"odcv": {
"da-7": 0,
"da-qwen-7": 0,
"delib-7": 0.4,
"delib-sonnet-7": 0.4
}
} | {
"training": {
"da-7": 0,
"da-qwen-7": 0,
"delib-7": 53,
"delib-sonnet-7": 0.4
},
"mask": {
"da-7": 0,
"da-qwen-7": 0,
"delib-7": 12.7,
"delib-sonnet-7": 0
},
"odcv": {
"da-7": 0,
"da-qwen-7": 0,
"delib-7": 0.4,
"delib-sonnet-7": 0
}
} | {
"da-7": {
"rows": 752,
"mentions_constitution_pct": 0.1,
"leak_check_pct": 0
},
"delib-7": {
"rows": 720,
"mentions_constitution_pct": 100,
"leak_check_pct": 87.8
},
"delib-sonnet-7": {
"rows": 751,
"mentions_constitution_pct": 24.5,
"leak_check_pct": 0.5
},
"da-qwen-... | {
"da-7": {
"mask_honesty": 74.2,
"odcv_mr_pct": 8.3
},
"delib-sonnet-7": {
"mask_honesty": 72.3,
"odcv_mr_pct": 17.1
},
"da-qwen-7": {
"mask_honesty": 81.5,
"odcv_mr_pct": 17.9
},
"delib-7": {
"mask_honesty": 69.3,
"odcv_mr_pct": 29.2
}
} | {
"da-7": "dougalldeepmind/2026-09-14-da-synth",
"delib-7": "dougalldeepmind/2026-09-16-delib-synth",
"delib-sonnet-7": "dougalldeepmind/2026-09-16-delib-sonnet-synth",
"da-qwen-7": "dougalldeepmind/2026-09-16-da-qwen-synth"
} | {
"da-7": "dougalldeepmind/2026-09-15-qwen36-0-da-7",
"delib-7": "dougalldeepmind/2026-09-16-qwen36-0-delib-7",
"delib-sonnet-7": "dougalldeepmind/2026-09-16-qwen36-0-delib-sonnet-7",
"da-qwen-7": "dougalldeepmind/2026-09-17-qwen36-0-da-qwen-7"
} |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
constitution references in reasoning traces, training corpus vs eval time (MASK, ODCV), for the four constitutional-SFT arms — Callum 2026-09-14: 'have a look at the inner thoughts of the trained models on these evals, and see whether they reference the constitution'
| field | value |
|---|---|
experiment |
constitution references in reasoning traces, training corpus vs eval time (MASK, ODCV), for the four constitutional-SFT arms — Callum 2026-09-14: 'have a look at the inner thoughts of the trained models on these evals, and see whether they reference the constitution' |
date_generated |
2026-09-17 |
constitution |
constitutions/abridged/constitution.md |
source_repo |
https://github.com/Matthew-Bozoukov/Lessons_from_constituitional_AFT @ 58b0226db97201bc912ad97962b855ff43631dab |
models |
{'classifier': 'google/gemini-3-flash-preview (temperature 0)', 'arms': {'da-7': 'dougalldeepmind/2026-09-15-qwen36-0-da-7', 'delib-7': 'dougalldeepmind/2026-09-16-qwen36-0-delib-7', 'delib-sonnet-7': 'dougalldeepmind/2026-09-16-qwen36-0-delib-sonnet-7', 'da-qwen-7': 'dougalldeepmind/2026-09-17-qwen36-0-da-qwen-7'}} |
generation_config |
{'sampling': 'temperature 0.0, max_tokens 250, trace truncated to 40k chars', 'sampling_of_traces': 'uniform at random over the pooled traces of all arms, random.Random(0); 1,000 MASK + 1,000 ODCV; training corpora labelled in full', 'schemes': {'reference': 'none |
schema |
labels.csv: scheme, eval (training |
provenance |
uv run python scratch/classify_constitution_refs.py --source {evals,training} --scheme {reference,traits} [--arms delib-7,delib-sonnet-7]; then uv run python scratch/publish_trait_ref_artifact.py |
- Downloads last month
- 52