chess-coach-32b-v6-distill

Chess Coach 32B v6-distill (behavior-in-weights research adapter)

A LoRA adapter that distills the tier-selection rule into the model WEIGHTS and is evaluated with NO grounding (no engine and no Maia in the prompt). This is the harder behavior-in-weights research result, not the shipped model: the live-served coach is chess-coach-32b-v6-dpo2, the best-DPO refinement of the chess-coach-32b-v4-qlora SFT base, served grounded.

Model family and roles

• chess-coach-32b-v4-qlora: the SFT base, the full-reproducible-eval reference, and the fallback endpoint.
• chess-coach-32b-v6-dpo2: the live-served coach on the demo (Modal chess-coach-v6dpo2-4bit-maia, tuned:true), the best-DPO refinement of the v4 SFT base.
• chess-coach-32b-v6-dpo: the earlier DPO adapter, superseded by v6-dpo2.
• chess-coach-32b-v6-distill: the behavior-in-weights research adapter, evaluated with no grounding (0.325 tier-policy match).

What it is

• Base: unsloth/Qwen3-32B-unsloth-bnb-4bit (the same 4-bit base as v4).
• Objective: SFT on the no-grounding move-review format, with targets drawn from the deep-verified engine-best and canonical tier moves in chess-coach-v6. The goal is to test whether the tier policy can be recovered from the weights alone when the engine and human-move grounding are removed from the prompt.

Honest result (corrected v6 benchmark, 120 held-out TEST, no grounding)

Re-scored in one controlled session against the deep-verified v6 labels, same strict extractor as the shipped v4 eval (RESULTS_STAGE4_CORRECTED.md). Both rows are scored WITHOUT grounding:

Stripped of grounding the untuned base essentially cannot coach: it names a move only 25% of the time and fabricates illegal or unsound moves without a sound list. The distilled adapter recovers the tier rule from its weights: tier-policy match 0.325 (+0.303, about 15x), names-a-move 0.983, move-soundness 0.653, distinct 0.461.

Honest limits

• Advanced-tier limit: per-tier tier-policy match is beginner 0.358, intermediate 0.400, advanced 0.217. The advanced target is the engine-best move, and reproducing it from the weights alone, without the engine grounding the condition removes, is genuinely the hardest, so advanced is the weakest tier.
• Grounding-free soundness (0.653) trails the deployable grounded soundness (about 0.98). This is a behavior-in-weights proof, not a claim that grounding is unnecessary in production.
• Tier-policy match is agreement with the project's select_tier_move rule, a learnability metric, not certified best teaching.

Use

Load the base 4-bit model and apply this adapter with PEFT (the tokenizer is the base Qwen3 tokenizer). Unlike the shipped grounded coach, this adapter is evaluated with the no-grounding prompt format described above.

Links

License Apache-2.0, inherited from Qwen3-32B. The training data is CC-BY-NC-4.0.