Cricket Stats Qwen3.5 0.8B (GGUF)

Cricket Stats is a small on-device model that turns verified cricket statistics into short, plain answers. It powers the optional "AI on this phone" answers in the StatYaar Android app, where it runs fully offline with llama.cpp.

Version: v6. Trained on about 12,000 realistic, messy questions built from the app's own prompts.

What it does

• Rewrites facts it is given (a player's saved stats) into a 1-3 sentence answer.
• Copies numbers exactly; it is not a source of cricket facts on its own.
• Says when a stat is not available, and declines predictions, betting, medical and opinion questions.

In the app, every number in a model answer is checked against the saved stats; if any number does not match, the app shows its own plain answer instead. Do not use this model as a source of facts without the same kind of check.

Files

Training

• Base model: Qwen/Qwen3.5-0.8B (Apache-2.0), LoRA fine-tune with MLX on Apple Silicon, merged and quantised to Q4_K_M with llama.cpp.
• Data: synthetic examples generated by code from invented player profiles (no real users' data, no personal data), plus (v5 only) filtered coaching Q&A from sahilml/cricket (Apache-2.0).
• Held-out eval (949 realistic questions, never seen in training): 99.1% of answers used only numbers present in the facts; 69.5% passed the app's full answer checker (stock base model: 54.1%).

Limitations

• Small model: phrasing can be plain or repetitive, and it may still invent a number, which is why answers must be checked against the facts.
• English first; Hindi and Hinglish are weaker.
• Not affiliated with or endorsed by CricHeroes or any cricket board.

Licence and attribution

Apache License 2.0 (see LICENSE). Based on Qwen3.5-0.8B by the Qwen team, Copyright 2026 Alibaba Cloud, licensed under Apache-2.0. Modified: LoRA fine-tuned on synthetic cricket-statistics answers, merged and quantised to GGUF Q4_K_M; text-only (vision and MTP weights removed).

Earlier versions (v5, still in this repo's history): contain information from Cricsheet (cricsheet.org) licensed under the Open Data Commons Attribution License (ODC-By 1.0); numbers were aggregated and anonymised. Also contain data from sahilml/cricket on Hugging Face, Apache License 2.0, modified (filtered, trimmed). The current v6 file uses only synthetic data generated by code.