| --- |
| library_name: transformers |
| license: other |
| license_name: lfm1.0 |
| license_link: LICENSE |
| base_model: LiquidAI/LFM2.5-230M-Base |
| pipeline_tag: text-generation |
| tags: |
| - chess |
| - lfm2 |
| - research |
| - transformers |
| --- |
| |
| # Chess language-model experiments |
|
|
| Five chess language-model checkpoints and evaluation summaries from experiments in [0xrushi/ChessLFM](https://github.com/0xrushi/ChessLFM). These checkpoints are provided for research and reproducibility. |
|
|
| **Model lineage:** [LiquidAI/LFM2.5-230M-Base](https://huggingface.co/LiquidAI/LFM2.5-230M-Base) → public [mlabonne/LFM2.5-230M-Chess](https://huggingface.co/mlabonne/LFM2.5-230M-Chess) → four experiments in this repository. The upstream ChessLFM checkpoint records LiquidAI's base model in its `surgery_meta.json`. Our four experimental checkpoints use the released ChessLFM weights as their **direct** starting point, so they descend from LiquidAI's base through ChessLFM. **They were not trained from random initialization or directly from the LiquidAI base weights.** The `released-baseline/` folder is the upstream ChessLFM reference, not a model trained by this project. |
|
|
| ## Checkpoints and observed results |
|
|
| | Rank by highest observed score | Folder | Checkpoint | Match score | W/D/L | Opening block | Descriptive Elo* | |
| | ---: | --- | --- | ---: | ---: | --- | ---: | |
| | 1 | `released-baseline/` | Released ChessLFM baseline | 17.5/32 | 12/11/9 | Later (16–31) | ~2033 | |
| | 2 | `branch-ranking-treatment/` | Branch-ranking treatment | 15.5/32 | 12/7/13 | Later (16–31) | ~1989 | |
| | 3 | `branch-ranking-control/` | Matched branch-ranking control | 13/32 | 5/16/11 | Later (16–31) | ~1934 | |
| | 4 (tie) | `self-play-refresh/` | Self-play refresh, round 1 | 12/32 | 9/6/17 | Initial (0–15) | ~1911 | |
| | 4 (tie) | `action-value-finetune/` | Action-value fine-tune, standard search | 12/32 | 10/4/18 | Initial (0–15) | ~1911 | |
|
|
| *The Elo column applies the logistic score transform relative to Stockfish configured at Elo 2000. It is descriptive only: 8–32 games per condition gives substantial uncertainty. Opening blocks differ, so ranks across blocks are only an ordering of the highest observed scores, not a controlled cross-model comparison. In the later block, the paired bootstrap interval for branch-ranking treatment versus control was [−4.0, +9.5] points; it includes no advantage. See the detailed [experiment report](https://github.com/0xrushi/ChessLFM/blob/main/EXPERIMENT_REPORT_2026-10-01.md). |
| |
| ## What is included |
| |
| Each folder contains Transformers model weights, model configuration, the chess tokenizer, and `token_ids.json`. Training recovery states, optimizer files, raw datasets, and match journals are not included. Use the project code for the chess-specific 80-token prompt, legal-move masking, and depth-three search; these weights alone are not a complete chess engine. |
| |
| Example loading one checkpoint: |
| |
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| repo = "0xrushi/chesslfm-experiments" |
| subfolder = "released-baseline" |
| tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=subfolder) |
| model = AutoModelForCausalLM.from_pretrained(repo, subfolder=subfolder) |
| ``` |
| |
| ## License |
| |
| The model weights are derivatives of [LiquidAI/LFM2.5-230M-Base](https://huggingface.co/LiquidAI/LFM2.5-230M-Base) through the upstream ChessLFM checkpoint. They are distributed under the [LFM Open License v1.0](LICENSE), including its commercial-use threshold. The repository's training code is separately available under MIT in the GitHub project. |
| |
| ## References |
| |
| - Upstream chess model: [mlabonne/LFM2.5-230M-Chess](https://huggingface.co/mlabonne/LFM2.5-230M-Chess) |
| - Base model and license: [LiquidAI/LFM2.5-230M-Base](https://huggingface.co/LiquidAI/LFM2.5-230M-Base) |
| - ChessLFM article: [ChessLFM: 2000 Elo in 230M params](https://maximelabonne.substack.com/p/chesslfm-2000-elo-in-230m-params) |
| |
