chess-coach-32b-v6-dpo2

Chess Coach 32B v6-dpo2 (best DPO adapter, live-served coach)

The best-DPO refinement of the v4 SFT base and the live-served coach on the demo (Modal endpoint chess-coach-v6dpo2-4bit-maia, tuned:true). It is a stronger, tier-targeted successor to chess-coach-32b-v6-dpo, built on the chess-coach-32b-v4-qlora SFT base, which remains the full-reproducible-eval reference and the fallback endpoint.

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• chess-coach-32b-v6-dpo: the earlier DPO adapter, superseded by v6-dpo2.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).
• Init: a two-adapter TRL DPO setup initialized from the v4 SFT LoRA for BOTH the policy and the frozen reference, so it is a clean superset of v6-dpo with no DPO-on-DPO drift.
• Training: harder, tier-targeted, style-matched preference pairs. Beginner learns to prefer the human-findable move over the sharp engine move; advanced learns to prefer the sharpest engine-best move over the softer sound move; intermediate keeps the original v6 pairs to preserve its gain. Selected checkpoint step 200. Preference data from chess-coach-v6.

Honest result (corrected v6 benchmark, 120 held-out TEST, grounded)

Re-scored in one controlled session against the deep-verified v6 labels, same strict extractor as the v4 reference eval (RESULTS_STAGE4_CORRECTED.md):

Overall tier-policy match 0.892 is the best DPO result: +0.031 vs v4 and +0.011 vs v6-dpo. Honestly, the entire gain is the intermediate tier (0.842 vs v4 0.750, vs v6-dpo 0.808). Beginner (0.858) and advanced (0.975) are byte-identical to v4 and v6-dpo because both already sit at their ceiling under grounding, so this is a stronger v6-dpo, not a beginner or advanced breakthrough. Soundness (0.983) and distinct-moves (0.987) are unchanged and names-a-move is nominally higher (0.986). Format (0.925) is marginally under v4 (0.939), a 256-token-cap prose-length artifact, not a move or soundness regression.

Honest limits

• The gain is confined to the intermediate tier; beginner and advanced are ceilinged.
• Tier-policy match is agreement with the project's select_tier_move rule, a learnability metric, not certified best teaching.
• Grounding is required at inference. Without the per-tier Maia signal the three tiers collapse to a single move; the behavior does not live in the weights.

Use

Load the base 4-bit model and apply this adapter with PEFT (the tokenizer is the base Qwen3 tokenizer). Run it with engine grounding in the prompt, the same move-review format v4 was trained on.

Links

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