Modelpublic

Chess-Nut-Engine/chess-qwen35-0.8b-sft-abc-20260708

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes

Model Card

chess-qwen35-0.8b-sft-abc-20260708

Qwen3.5-0.8B (Gated-DeltaNet hybrid) fine-tuned on the chess-sft-corpus-4x curriculum, 2M examples per phase, chained A (foundation) -> B (understanding) -> C (planning) on 2x H100. Each phase_*/ folder is that phase's best checkpoint; phase_c is the final model.

Benchmark results (greedy, frozen 13k benchmark)

Tier	A	B	C
T1 perception	98.9%	99.0%	98.9%
T2 rules	88.0%	87.1%	87.5%
T3 tactics	-	54.1%	57.3%
T4 evaluation	-	88.0%	88.1%
T5 openings	-	51.3%	53.5%
T6 endgames	-	70.7%	70.7%
T7 planning	-	-	31.1%
WPD (depth 20)	-	-	0.435

Phase A passes all foundation gates. Known weaknesses (2026-07-08 review): in-check move generation (check-blindness), move selection (answer-first traces). Both are addressed in the v2 data revision; SDPO planned for move quality. Full writeup in the repo's docs/experiments/2026-07-08_abc_run/README.md.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(
    "Chess-Nut-Engine/chess-qwen35-0.8b-sft-abc-20260708",
    subfolder="phase_c", trust_remote_code=True)
t = AutoTokenizer.from_pretrained(
    "Chess-Nut-Engine/chess-qwen35-0.8b-sft-abc-20260708",
    subfolder="phase_c")

<move>/</move> are tokenizer special tokens (vocab resized +2). Planning prompts use the <think>...</think><move>uci</move> protocol.