Evangelinejy's picture
model_20m_16B card
6727e61 verified
|
Raw
History Blame Contribute Delete
1.5 kB
metadata
license: apache-2.0
tags:
  - chess
  - reinforcement-learning
  - grpo

model_20m_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_20m_16B and model_20m_16B.

parameters 20m
pretraining tokens 15,824,042,105 (15.8B)
checkpoints here 30 (steps 100–3000)
checkpoints saved by the run 60

Steps

100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1100, 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900, 2000, 2100, 2200, 2300, 2400, 2500, 2600, 2700, 2800, 2900, 3000

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_3000"
p = snapshot_download("Pre2Post-Chess-RL/Chess-RL-Models", allow_patterns=f"model_20m_16B/{step}/*") + f"/model_20m_16B/{step}"
model = AutoModelForCausalLM.from_pretrained(p, trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(p, trust_remote_code=True)