Datasetpublic

dougalldeepmind/2026-09-15-da-multiagent-7-mix

multi-agent principle-10 arm: the MSM Table 2 base blend scaled around a synthetic difficult-advice share written against principle 10 alone field value experiment multi-agent principle-10 arm: the MSM Table 2 base blend scaled around a synthetic difficult-advice share written against principle 10 alone — final training mixture (synthetic sources mixed in) date_generated 20260915 constitution… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-09-15-da-multiagent-7-mix.

sourceHugging Faceupdated 11d agoView on Hugging Face

0likes72downloads

Dataset Card

multi-agent principle-10 arm: the MSM Table 2 base blend scaled around a synthetic difficult-advice share written against principle 10 alone

field	value
`experiment`	multi-agent principle-10 arm: the MSM Table 2 base blend scaled around a synthetic difficult-advice share written against principle 10 alone — final training mixture (synthetic sources mixed in)
`date_generated`	20260915
`constitution`	constitutions/experimental/claudedistilled10principlesmultiagent/constitution.md
`source_repo`	https://github.com/Matthew-Bozoukov/Lessonsfromconstituitional_AFT.git @ 6214d8a0d0b882dd92f2d61b34d91293c5c32139
`models`	base-blend reasoning traces: on-policy qwen/qwen3.6-27b (family qwen36, 1135 turns, judged by google/gemini-3-flash-preview)
`generation_config`	{"seed": 0, "maxseqlen": 8192, "budgettokenizer": "Qwen/Qwen3.6-27B", "basemixture": {"repo": "dougalldeepmind/2026-09-08-nosynth-mix", "file": "mixture.jsonl", "revision": "7e991f58e86eff0b0a9f15a54ebeddfffb5b14dd"}, "reasoningtraces": {"model": "qwen/qwen3.6-27b", "family": "qwen36", "judge": "google/gemini-3-flash-preview", "sources": ["tulu3if", "selfossinstruct", "lima"], "fraction": 0.5, "seed": 0, "rows": null, "turns": 1135, "provenance": "enrichmentreport.json of dougalldeepmind/2026-09-08-nosynth-mix@7e991f58e86e (scratch/reasoningbackfill, pre-record)", "inherited_from": {"repo": "dougalldeepmind/2026-09-08-nosynth-mix", "revision": "7e991f58e86eff0b0a9f15a54ebeddfffb5b14dd"}}}
`schema`	mixture.jsonl + mixturestats.json. jsonl rows {messages: [{role, content, reasoningcontent?, toolcalls?}], tools?, source, supervise?} — model-agnostic interchange; rendered with the training family's chat template at train time (src/modelprofile.py render_chat), tools included
`provenance`	uv run scripts/data/mixture/build_mixture.py --config configs/data/mixture/da-multiagent.yaml