Downloads · 30 days
17
36% of all-time downloads
rukh-qwen3-pgn-qlora
chorcat/rukh-qwen3-pgn-qlora
rukh-qwen3-pgn-qlora is a machine learning model from chorcat. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
A QLoRA adapter that fine-tunes Qwen/Qwen3-0.6B on chess games written as PGN text. It exists to answer one question with a number: is a general language model, fine-tuned on the same games, a substitute for a model b…
All-time downloads
47
Public
Repo size
20.6 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.json11.4 MB · 55%
At a glance
- Library
- peft
- License
- apache-2.0
- Access
- Public
- Created
- Sep 20, 2026
- Updated
- Sep 21, 2026
- SHA
- 44bc01b3
Base models
Model card
From the Hugging Face model README
chorcat/rukh-qwen3-pgn-qlora
A QLoRA adapter that fine-tunes Qwen/Qwen3-0.6B on
chess games written as PGN text. It exists to answer one question with a number: is a general
language model, fine-tuned on the same games, a substitute for a model built for the job?
It is part of Rukh, a course that builds a chess language model end to end. The model it is compared against, and how: https://lab.rukh.borjaglez.com.
How it was trained
| Base model | Qwen/Qwen3-0.6B (600,637,440 parameters) |
| Precision | 4-bit NF4 (QLoRA) |
Rank r | 16, alpha 32 |
| Adapted matrices | k_proj, o_proj, q_proj, v_proj |
| Trainable parameters | 4,587,520 (0.764 %) |
| Training samples | 100,000 games |
| Steps | 1,500 |
| Weights on the device | 851 MB |
| Peak memory while training | 2361 MB |
The games are the same ones the project's own decoder trains on, from the same month, held out against the same validation split. Each is rendered as the movetext a human would paste into a chess site, with the two ratings as PGN tags:
[WhiteElo "2015"] [BlackElo "2028"]
1. e4 c6 2. d4 d5 3. exd5 cxd5 4. Bd3 Nf6 ... 1-0
Results
Measured with the project's own harness -- the same validation positions, the same puzzle split and the same Stockfish ladder as every other stage -- so the two sides differ in their representation and in nothing else.
| Metric | Value |
|---|---|
| Legal moves written, no mask | 62.50 % |
| Top-1 next move | 12.5 % |
| Puzzles solved | 0.9 % |
| Estimated Elo | < 807 (one-sided 95 % bound; every game lost) |
How its answers failed
A model whose vocabulary is the set of legal moves can only be wrong one way: a legal move in the wrong position. A model writing SAN has three more.
| Outcome | Share |
|---|---|
| legal | 62.50 % |
| illegal here | 33.70 % |
| not a move at all | 2.90 % |
| ambiguous SAN | 0.90 % |
| nothing written | 0.00 % |
In the course
- Built in M4 · Fine-tuning: los labs, the lesson that runs every command behind this repository.
- Measured in the single results table as the stage
qwen3-pgn-qlora: every stage of the course, the same suite, the same day. - Bring it to the paths the configs read:
uv run rukh pull qwen3-pgn-qlora.
Licence
apache-2.0.
- Library
- peft
- License
- apache-2.0
- Created
- Sep 20, 2026
- Updated
- Sep 21, 2026