Qwen3-8B Backgammon (GGUF)

Files

Usage (llama.cpp)

Training

• Base: Qwen/Qwen3-8B (Apache 2.0)
• Method: QLoRA (r=64, alpha=64) via Unsloth on Kaggle T4. Two stages: continued pretraining (~870k tokens of backgammon literature, 1 epoch, lr 5e-5) then SFT (5,000 instruction pairs, 74% Japanese / 26% English, 3 epochs, lr 2e-4, train-on-responses-only). Final SFT train loss 1.24.
• Data: a curated backgammon corpus assembled from classic strategy literature, Usenet (rec.games.backgammon) archives, engine documentation, match-equity tables and probability references, plus programmatically generated worked examples (Thorp count, Janowski take points, MET lookups). The corpus itself is not redistributed.

Behaviour & limitations

• Compared to the stock model, it correctly grounds backgammon terminology (e.g. identifies the Thorp count as Ed Thorp's race cube formula rather than guessing) and answers in a dense, no-fluff style in Japanese or English.
• It does not reliably reproduce exact formulas or table values. Details of multi-step formulas (Thorp count adjustments, MET cell values, etc.) can be confabulated. Treat numeric output as unverified; use a reference or RAG for exact values.
• 8B-class general reasoning limits apply. Not a playing engine — it explains concepts; it does not compute best moves.

License & attribution