Beyond the Move / Além do Lance

A rank-8 LoRA adapter fine-tuned on CPU for a Brazilian Portuguese chess practice experiment. Use it with Qwen/Qwen2.5-0.5B-Instruct at revision 7ae557604adf67be50417f59c2c2f167def9a775. The base weights are not bundled.

The adapter proposes a coaching focus and a short question given a natural-language rationale and three board flags. It is not a chess engine and receives no solution line. See the project for the exact prompt, notation guard, UI and reproduction commands.

Training and results

Chess positions: Lichess puzzle export, CC0 1.0. Coaching rationale and target pairs are original synthetic examples. 48/8/12 source position groups for training/validation/test with disjoint phrasing pools. One of two trials selected on validation; test run once. 540,672 LoRA parameters, 180 CPU steps, seed 42 and about 434 seconds.

On the 60 synthetic held-out examples Focus accuracy Macro F1 Invalid output
Base under the same prompt 0/60 0.000 60/60
Keyword rules 12/60 0.067 0/60
LoRA adapter 31/60 0.483 0/60

Strict five-intent group accuracy was 0/12. The adapter frequently confused defense with attack and capture with a trade. Explicit move notation was not observed in this small test; hidden spoilers, factual errors, prompt adherence and usefulness on real learner language are not guaranteed. The base result includes format failures on one prompt. This is a small instruction-following experiment, not an Elo or teaching-effectiveness result.

metrics.json contains the confusion counts. training-manifest.json pins the data, training code and artifact digest. Weights use safetensors.

Downloads last month
12
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for pedrojb04/beyond-the-move

Adapter
(1,075)
this model