QLoRA adapter that teaches a large open model one chess-coaching behavior: given a position and the student's rating tier, select the tier-appropriate, instructive move and tag it with a short principle (for example, Nf3 - develop toward the center). This v4 adapter is the SFT base, the full-reproducible-eval reference, and the fallback endpoint of the chess-instructor-llm project ("The Analysis Room"). The live-served coach is now its best-DPO refinement chess-coach-32b-v6-dpo2.

Base model: unsloth/Qwen3-32B-unsloth-bnb-4bit — a 4-bit bitsandbytes quantization of Qwen/Qwen3-32B.
Method: QLoRA / LoRA, rank r=32, lora_alpha=32, dropout 0.0, applied to all attention and MLP projections (q,k,v,o,gate,up,down), trained with Unsloth + TRL (PEFT 0.19.1).
Adapter size: adapter_model.safetensors ≈ 1.07 GB.

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chess-coach-32b-v6-dpo: the earlier DPO adapter, superseded by v6-dpo2.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).

The one trained behavior

Given a chess position, the student's rating tier (beginner ~1000-1200, intermediate ~1300-1600, advanced ~1700-2000), the move the student played, and engine grounding (a Stockfish sound-move pool with short lines plus Maia human-move likelihoods at that tier), the model's single trained job is to choose the tier-appropriate instructive move and render it as move - short principle. That choice is graded deterministically, with no model judge, on three clauses a stranger with an engine and a human-move model can check:

sound — not a blunder (engine cp-loss below 250);
tier-appropriate — matches the canonical tier move (the most human-findable sound move for a beginner, the engine's sharpest line for an advanced player);
distinct across levels — a beginner and an advanced player are not handed the same move on a differentiating position.

The English explanation ("prose") is not the trained behavior. It is an optional display layer that can be rendered by engine templates or a frontier model and is verified separately by a non-LLM checker. See Limitations.

Results (base vs. tuned)

Held-out validation slice: 120 positions x 3 tiers. Grounding is held byte-identical between rows, so the only thing that changes between base and tuned is the weights. Move selection is scored deterministically against the canonical tier move, using the STRICT any-legal move extractor (accept any clearly-named legal move; no in-pool backfill) — the single canonical parser shared with the grand eval.

Contender (same grounding)	tier-fit ↑	coherence violation ↓	distinct moves / level ↑
Qwen3-32B base (untuned)	0.347	0.500	0.290
v4 (this adapter)	0.77	0.140	0.73
best frontier on tier-fit (Gemini 3.1 Pro)	0.553	0.292	0.210

distinct-moves-per-level = 73/100 canonical beginner!=advanced opportunities (a no-answer counts as a miss); the grand-eval table's 0.785 uses the 73/93 denominator = only the positions where v4 named both tier moves.

Note: "tier-fit" = "tier-policy exact match" (exact agreement with the preregistered select_tier_move rule, a project rule, not validated pedagogy).

Moat (head-to-head vs. the best frontier). Of the 120 positions, v4 gives a distinct, sound, correctly-graded per-tier move on 67, and of those it diverges from the best frontier's move on 62. On those 62 diverging positions v4 wins 51, loses 5, ties 6 on the moat (tier-fit then soundness) against the best frontier at each position. That 51-5-6 is a selection-conditioned subset (it conditions on v4 first giving a distinct, sound, correctly-graded move), so it overstates a raw win rate; the honest unbiased, reproducible head-to-head over all 92 positions where v4 diverges from the best frontier is 56-24-12 (56-24-40 over all 120; recomputed from the committed raw/greedy gens by scripts/reproduce_v4.py), and the primary unbiased comparison is the all-scenario tier-fit 0.767 vs 0.553. The frontier loses this axis because it hands the engine's single best move to every level, repeating one move across the three tiers about 77% of the time.

The result is off-consensus: a well-engineered prompt on the same base does not close the tier-fit gap (measured at 1.7B, 4B, and 32B), because tier-appropriate move selection is added by data, not by prompting.

How to use

The adapter files (LoRA weights, tokenizer, chat template) are at the repo root. Load the exact 4-bit base and attach the adapter with PEFT:

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "unsloth/Qwen3-32B-unsloth-bnb-4bit"   # 4-bit bnb quant of Qwen/Qwen3-32B (requires bitsandbytes)
ADAPTER = "khoilamalphaai/chess-coach-32b-v4-qlora"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
base = AutoModelForCausalLM.from_pretrained(BASE, device_map="auto", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, ADAPTER).eval()

messages = [
    {"role": "system", "content": "You are a chess coach doing move review. Recommend one sound, tier-appropriate move and tag it with a short principle."},
    {"role": "user", "content": "Student tier: beginner.\n[Engine-grounded prompt: FEN, played move, Stockfish sound-move pool + short lines, Maia likelihoods.]"},
]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Notes:

The base is a pre-quantized 4-bit checkpoint, so bitsandbytes is required and no separate BitsAndBytesConfig is needed. You can also load it in one step with unsloth's FastLanguageModel.from_pretrained pointed at this adapter repo.
The model is trained for the grounded move-review format above. Run it with engine grounding in the prompt; do not expect reliable chess reasoning ungrounded.

Limitations

Prose is weaker than smaller checkpoints, by design. v4 is the strongest checkpoint on the move and simultaneously the weakest recent one on prose: its blinded instructiveness grade is 4.67 (below the 4B tune at 5.32 and the prior v3 at 6.35), and about 40% of its raw drafts trip the prose faithfulness check before the shipped gate. This is a non-issue for the graded claim, which is the move: render prose with engine templates or a frontier model on top of the chosen move and gate it with the non-LLM verifier.
Not the local form factor. This 32B adapter is the reference that shows scale buys a still-stronger move (and beats the frontier on the moat). The on-spec production target is a small ~4B local coach; this repo is for the strongest move, not for laptop inference.
Grounding-dependent. Move truth comes from the engine + tier rule; the adapter is the last-mile chooser. Truthfulness of any rendered prose relies on the inference-time verifier, not the weights.

Links

Resource	Link
Live-served coach (best DPO)	chess-coach-32b-v6-dpo2
Code / GitHub repo	Alpha-AI-Engineering-Khoi/chess-instructor-llm
Training dataset (SFT)	chess-coach-move-review (`v4` config)
Evaluation data	chess-coach-grand-eval
Demo	chess-coach-studio
Sibling model (1.7B / v3, MLX)	qwen3-1.7b-chess-coach-mlx

Training details

Trained on the v4 split of khoilamalphaai/chess-coach-move-review: distilled, engine-grounded, tier-contrastive move reviews with a hard faithfulness filter (0% false move labels).
QLoRA on unsloth/Qwen3-32B-unsloth-bnb-4bit, r=32, alpha=32, dropout 0.0, LoRA on q,k,v,o,gate,up,down, task type CAUSAL_LM (Unsloth + TRL, PEFT 0.19.1).
License: Apache-2.0, inherited from Qwen3-32B.