Downloads · 30 days
0
chess-gpt-demo-ngram
peterwooden/chess-gpt-demo-ngram
chess-gpt-demo-ngram is a machine learning model from peterwooden. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This is the first playable teaching baseline from peterwooden/chess-gpt. It is a count-based backoff n-gram model, not a neural network or Transformer, and exists to establish a reproducible floor before neural experi…
Access
Public
Updated Jul 26, 2026
Repo size
1.4 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.json6.7 MB · 83%
At a glance
- Access
- Public
- Created
- Jul 25, 2026
- Updated
- Jul 26, 2026
- SHA
- 5ec1a203
Model card
From the Hugging Face model README
Chess GPT demo SAN n-gram
This is the first playable teaching baseline from peterwooden/chess-gpt. It is a count-based backoff n-gram model, not a neural network or Transformer, and exists to establish a reproducible floor before neural experiments begin. It predates the agreed 2026 tournament corpus and is not tournament-eligible.
Play it
Open the Chess GPT browser arena and load this pinned package reference:
peterwooden/chess-gpt-demo-ngram@bea221167728c33f0a5df54051cd27717cae6586
The browser verifies every declared file before running the adapter locally in a dedicated Worker.
Interface
- Input: the complete game history as standard algebraic notation (SAN) moves.
- Output: exactly one legal SAN move.
- Policy: use the most frequent legal continuation of the last two moves; back off to a one-move context, then side-to-move frequencies, then a deterministic legal fallback.
- Package:
chess-gpt-package-v1with a self-containedbrowser/entry.js, canonicalbrowser/model.json, and hashedbrowser/manifest.json.
Training and evaluation
- Dataset:
shazmate/lichess-chess-tokens - Dataset revision:
cb90f1bb2eab0b905e84e14f2d1d24ec5f9d1d94 - Games processed: 10,000
- Split: deterministic whole-game split; 8,990 training games and 1,010 validation games
- Validation top-1 accuracy: 22.41%
- Validation legal-move rate: 100%
- Compressed checkpoint: 1,412,649 bytes
- Canonical learned state: 6,687,169 bytes
- Checkpoint SHA-256:
c143646b163c8c91c39bc654bf9a57d0a142bf010ec7b9c90971248616103f1c - Canonical package: 6,689,698 bytes
- Manifest SHA-256:
fc1160f487f315036fc320a7e7ced5a2e8cf9872345318a2959f8f727b949702
The full machine-readable record is in metrics.json. The experiment specification and implementation live in the source repository as experiment 0001-basic-san-ngram.
Limitations
This model recognizes short statistical continuations; it does not evaluate positions, search variations, or understand long-range plans. Validation next-move accuracy is not a claim about playing strength. Expect weak and sometimes strange chess—the point is to make the first floor observable and beatable.
- Created
- Jul 25, 2026
- Updated
- Jul 26, 2026