Second-1.0-2B: Grounded Chess Explanation Model

Second is an open-source line of small language models that explain chess clearly. It is based on Qwen/Qwen3.5-2B. It turns engine facts into clear human coaching explanations for chess players.

Stockfish supplies the ground truth. The model does not calculate; it explains.

• Organization: Second Chess
• Repository: https://github.com/kleinebossie/second
• Model Family: Second 1.0 (2B, 0.8B, 4B, 9B)
• Browser Quantization: second-chess/Second-1.0-2B-GGUF

---

Model Description

Chess engines show evaluation bars and lines of engine symbols (+2.4, 14. Nxd4 exd4 15. Qh5+), but amateur players cannot understand why a move works or fails. Second bridges this gap by converting deterministic chess engine facts (tactical motifs, hanging pieces, refutations, material balance, and king safety) into clear, instructive prose.

Second-1.0-2B supports two target audiences:

1. beginner (ratings under ~1200): Plain English, shorter sentences, no complex jargon (no "prophylaxis", "zwischenzug", or "outpost").
2. club (ratings ~1200–2000): Concrete tactical themes, refutation reasons, and strategic ideas.

---

Input Prompt Contract

Second uses the standard Qwen chat template. The user prompt follows a strict deterministic schema:

task: explain_move
level: club
detail: short

FEN: r1bqkbnr/1p1n1ppp/p2p4/2pPp1BQ/B3P3/8/PPP2PPP/RN2K1NR b KQkq - 1 7
Black to move, move 7.
Move played: Nf6.
Verdict: Blunder.
Black's winning chances: 89% → 57%.
Best move: Qa5+.
Best line: Qa5+ Nc3 b5 Bb3 Ngf6 Qe2.
Facts:
- Nf6 attacks the queen on h5.
- After Nf6, White's best reply is Bxd7+.
- The queen on h5 is not defended.
- White has a check: Bxd7+.
- White can capture the knight on f6 with Bxf6.
- Black threatens Nxh5.
- Material: equal.
- Phase: opening.

Example Model Output

The move Nf6 is a blunder because it ignores the immediate tactical refutation. While attacking the queen on h5 looks tempting, White responds with Bxd7+, forcing Black to lose material or suffer a devastating counterattack. The best move is Qa5+, which maintains pressure and avoids the loss of the queen. This check keeps the position balanced and prevents White from gaining a decisive advantage.

---

Stage T6 Frozen Evaluation Benchmarks

Second-1.0-2B was evaluated on the frozen 300-item test set (eval/testset-1.0.jsonl) with 9 deterministic verifiers and the debiased Gemini Flash judge:

Metric	Base Qwen 2B	Second v1 (grpo-1.0-2b)	Second-1.0-2B (Release)
Deterministic Hard-Fail Rate	100.0%	7.7%	4.67% (14 / 300)
Format Pass Rate	82.3%	100.0%	100.00%
Target Length Compliance	21.0%	18.7%	99.67% (62.0 avg words)
First-Person Diary Voice	12.0%	51.3%	0.00%
Key-Point Soft Score	0.2100	0.3860	0.9790
Better-Move Soft Score	0.0800	0.1560	1.0000
Aggressive Terminology Over-Refusal	15.0%	0.0%	0.00% (0 / 20)
Pairwise Win Rate vs Base Qwen 2B	—	—	100.0% (86.2% LC-WR)
Pairwise Win Rate vs Second v1	—	—	100.0% (97.5% LC-WR)
Pairwise Win Rate vs Gemini Flash	—	—	91.0% (74.5% LC-WR)

---

Training Recipe

Second-1.0-2B was trained using a two-stage post-training pipeline on an NVIDIA L4 GPU:

1. Stage T3 (Engine-Grounded Distillation Warm-up SFT-A1):
   • Trained with LoRA (r=32, alpha=64, all linear layers) on 4,457 verified items (data/processed/sft_v2_train.jsonl).
   • Combined Apache 2.0 teacher synthetic explanations (Qwen2.5-7B-Instruct) filtered by 9 rule checks with curated gold human commentary.
2. Stage T4 (GRPO RL Warm-up RL-B2):
   • Trained for 150 steps using DAPO (Direct Advantage Policy Optimization, beta=0.02) across 5,000 difficulty-filtered prompts (data/processed/rl_prompts_v2.jsonl).
   • Reward v2: Graded rule check penalties, multiplicative length rescaling, and anti-diary voice shaping. No LLM judge inside the RL loop.

---

Limitations

• Engine Dependency: Stockfish supplies the ground truth. If the engine analysis depth is insufficient, the facts may miss deep multi-ply variations.
• Language: English language only.
• Calculation: The model does not compute chess variations beyond the engine-provided facts and best lines.
• Ratings: Targeted at beginner (under 1200) and club (1200–2000) players. Not designed for master-level opening novelty analysis.

---

Fair Play and Integrity Notice

Second is designed exclusively for post-game review and tactical training. The official browser extension fails closed:

• Second turns completely OFF during active live games, blitz games, puzzles, puzzle storms, and live tournaments.
• Second only activates on finished games (/analysis, study boards, or ended game URLs).
• Do not use this model or its outputs for live game assistance.

---

License

Apache License 2.0. See LICENSE for details.