Karvonen ChessGPT 50M MLX BF16

This repository contains the local MLX/bfloat16 conversion of Adam Karvonen's 16-layer ChessGPT checkpoint used by the ChessGPT negative-control runs in the ICLManyReplication artifact for "Many Next-Token Predictors are In-Context Learners".

The local config records the source as:

source repo: adamkarvonen/chess_llms
source checkpoint: lichess_16layers_ckpt_no_optimizer.pt
source dataset: 7GB Lichess Dataset
source iteration: 600000
parameters: 50,888,704

The upload includes the converted model.safetensors, tokenizer files, and the custom chess_gpt.py model implementation used by the replication harness.

Intended Use

This checkpoint is intended as a drop-in artifact for reproducing the ChessGPT negative-control row in ICLManyReplication. Inputs are character-level PGN; the source model notes that prompts should be prefixed with ;, the training delimiter between games.

Downloads last month: 236

Safetensors

Model size

50.9M params

Tensor type

BF16

MLX

Hardware compatibility

Quantized

Model tree for N8Programs/karvonen-chessgpt-50m-mlx-bf16

Base model

adamkarvonen/chess_llms

Finetuned

(1)

this model