---
license: apache-2.0
tags: [chess, reinforcement-learning, grpo]
---

# model_20m_53B — 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_53B`](https://huggingface.co/pavelslab-nyu/Chess-Pretrain-Models/tree/main/model_20m_53B) and
[`model_20m_53B`](https://huggingface.co/pavelslab-nyu/Chess-SFT-Models/tree/main/model_20m_53B).

| | |
|---|---|
| parameters | 20m |
| pretraining tokens | 52,746,807,017 (52.7B) |
| checkpoints here | 32 (steps 100–5000) |
| checkpoints saved by the run | 162 |

## Steps

`100`, `200`, `300`, `400`, `500`, `600`, `700`, `800`, `900`, `1000`, `1100`, `1200`, `1300`, `1400`, `1500`, `1600`, `1800`, `2200`, `2400`, `2600`, `2800`, `3000`, `3200`, `3400`, `3600`, `3800`, `4000`, `4200`, `4400`, `4600`, `4800`, `5000`

## 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**:

```python
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

step = "global_step_5000"
p = snapshot_download("Pre2Post-Chess-RL/Chess-RL-Models", allow_patterns=f"model_20m_53B/{step}/*") + f"/model_20m_53B/{step}"
model = AutoModelForCausalLM.from_pretrained(p, trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(p, trust_remote_code=True)
```