Hadron is an LLM distillation framework built around NousResearch's AutoReason tournament refinement: a single teacher model produces distillation labels that are measurably better than its own single-shot output. Each prompt is answered, critiqued, adversarially revised, synthesized, and ranked by a blind Borda panel until "do nothing" wins twice. Full reasoning traces are captured per role for process-supervision fine-tuning.
The core AutoReasonedGeneration Task runs each prompt through five fresh-context agent roles, each solving a different problem:
- Teacher — writes the incumbent draft A. Defines the quality ceiling; should be the strongest model you can afford.
- Critic — finds concrete, quotable flaws in A or replies exactly
NO FLAWS. Needs discrimination, not creativity. Anti-hallucination directives keep it honest. - Author B — adversarial revision. Rewrites A to address the critique without padding or scope creep.
- Synthesizer — conservative synthesis AB. Minimum repair: keep what worked, fix only what was flagged.
- Judges (N) — blind Borda panel. Each judge sees A, B, AB under randomised labels and returns one ranking line. Panel size beats panel quality — 7 noisy judges outperform 3 careful ones.
If A defends twice in a row the tournament converges. Otherwise the winner is promoted and the loop repeats. Every call is rate-limited and retry-wrapped: 429s, 503s, Metal working-set rejections, and httpx timeouts all backoff-and-retry instead of dropping votes.
A 69-sample pilot on Gemma 4 26B-A4B (local vMLX, thinking on) produced rich traces at ~8 min/sample with zero unrecovered failures. Blind external-judge evals on a prior pilot picked AutoReason over single-shot baseline on all 3 randomised runs.
- SFT fine-tuning — train students on the refined
generationfield (higher-quality targets than raw teacher output). - DPO / preference training — every tournament iteration produces winner/loser pairs with Borda scores for free.
- Chain-of-thought distillation — per-role reasoning traces teach students how to think, not just what to say.
- Critic-model training —
(draft, critique, revision)triples for self-correcting students. - Agentic specialisation — paired hand-authored agentic prompts cover tool use, planning, reflection, error recovery, and multi-agent coordination.
Vanilla distillation is one teacher pass per prompt — whatever the model says first is what the student learns. Best-of-N sampling picks the best of several drafts but still treats each draft as atomic. AutoReason refuses both frames: the teacher's first answer is a starting point, its own critic is invited to stress-test it, and "no change needed" is a first-class winning move rather than a default assumption. The same model is three different workers, and self-evaluation is almost always stronger than self-generation — that's the lever we pull.
- Python 3.11, asyncio, pydantic v2
- Built-in pipeline runtime (async, pydantic v2, DAG-based Step/Task graph)
- AutoReason tournament (A / B / AB + blind Borda) adapted from the NousResearch paper
- OpenAI async SDK — OpenRouter, local vMLX, any OpenAI-compatible endpoint
- MLX-LM for on-device Apple Silicon inference (Gemma 4, Qwen3)
- pytest + pytest-asyncio (61 unit + integration tests)
- Hugging Face datasets (
ianncity/General-Distillation-Prompts-1M)
Active