Evangelinejy's picture
cards: drop alpha/beta/compute/run-id
f75c74f verified
|
Raw
History Blame Contribute Delete
1.3 kB
metadata
license: apache-2.0
tags:
  - chess
  - reinforcement-learning
  - grpo

model_680m_16B — RL (GRPO) checkpoints

RL post-training trajectory for the chess pre-to-post compute-allocation study. The pretraining base and SFT init for this model are model_680m_16B and model_680m_16B.

parameters 680m
pretraining tokens 15,994,915,000 (16B)
checkpoints here 4 (steps 1200–1900)
checkpoints saved by the run 25

Steps

1200, 1400, 1700, 1900

Loading

Each global_step_N/ folder is self-contained. The models use a custom tokenizer (tokenizer.py), and the remote-code resolver ignores subfolder=, so download the folder first and load the local path:

from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

step = "global_step_1900"
p = snapshot_download("Pre2Post-Chess-RL/Chess-RL-Models", allow_patterns=f"model_680m_16B/{step}/*") + f"/model_680m_16B/{step}"
model = AutoModelForCausalLM.from_pretrained(p, trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(p, trust_remote_code=True)