Praxis Chess LoRA v1: the comparison arms
Research artifacts, not models to use. For explaining chess mistakes, use praxis-chess-reasoner-qwen3.5-2b-lora or the 4B.
The study asked one question: does a small model explain chess mistakes more truthfully when it is given structured evidence? Every model below is the same base and recipe as the released R3 adapters; only the input changes. They are published so that every row of the results table can be reproduced, not only the winning one.
| Folder | Base | Input it was trained and tested on | Claim precision, public test (95% CI) | Chain validity (95% CI) |
|---|---|---|---|---|
2b-r0/ | Qwen3.5-2B | R0: board and move | 34.8% (31.4–38.1) | 0.5% (0.1–1.1) |
2b-r1/ | Qwen3.5-2B | R1: + engine numbers | 56.9% (54.7–58.9) | 14.4% (12.1–16.8) |
2b-r2/ | Qwen3.5-2B | R2: + flat facts | 56.8% (54.6–58.8) | 12.6% (10.6–14.9) |
4b-r0/ | Qwen3.5-4B | R0: board and move | 32.6% (29.5–35.8) | 0.6% (0.1–1.2) |
| released 2B-R3 | Qwen3.5-2B | R3: the full evidence graph, with item IDs | 99.7% | 78.0% |
| released 4B-R3 | Qwen3.5-4B | R3 | 99.7% | 86.8% |
- R0: the position (FEN), the player's side and the move played.
- R1: R0 plus the engine's best move, the evaluation before and after, and the engine line after the played move.
- R2: R1 plus a flat list of what the move changed: what it attacks, pieces left loose or without a defender, lines opened or closed. No IDs.
- R3: R2's facts plus what only the graph holds: the threat before the move, the counterfactual after the best move, and how deep the engine had to look, as typed nodes with IDs that a claim can cite.
Each folder holds the adapter (adapter/), the merged q4_K_M GGUF the numbers
were measured with, its Modelfile, the training configuration and the
checkpoint-selection scores. Reproduce any row with the Praxis repo:
python reproduce.py run --model 2b-r0 # or 2b-r1, 2b-r2, 4b-r0
The five 2B-R3 ablations (each trained with one evidence component removed, for H5) were evaluated only on the author's private games, so they are not published here.
Training: Qwen3.5-2B (revision 15852e8c16360a2fea060d615a32b45270f8a8fc) or Qwen3.5-4B (revision
851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a), LoRA r = 16, alpha = 32, two epochs, learning rate
2e-4, answer tokens only, one NVIDIA A40. The released adapters' cards have the
full method, the results on the author's games and the limitations.
Licence: Apache-2.0, like the base models.