model_50m_23B/README.md

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

# model_50m_23B — 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_50m_23B`](https://huggingface.co/pavelslab-nyu/Chess-Pretrain-Models/tree/main/model_50m_23B) and
[`model_50m_23B`](https://huggingface.co/pavelslab-nyu/Chess-SFT-Models/tree/main/model_50m_23B).

| | |
|---|---|
| parameters | 50m |
| pretraining tokens | 22,910,846,607 (22.9B) |
| checkpoints here | 10 (steps 100–1000) |
| checkpoints saved by the run | 20 |

## Steps

`100`, `200`, `300`, `400`, `500`, `600`, `700`, `800`, `900`, `1000`

## 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_1000"
p = snapshot_download("Pre2Post-Chess-RL/Chess-RL-Models", allow_patterns=f"model_50m_23B/{step}/*") + f"/model_50m_23B/{step}"
model = AutoModelForCausalLM.from_pretrained(p, trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(p, trust_remote_code=True)
```