---
library_name: transformers
tags:
- policy-gradient
- adaptive-group-size
---

# Adaptive group-size policy checkpoints

Final fixed-group and adaptive policy checkpoints for GSM8K, mathematical
reasoning and code generation. This is a multi-model archive: choose a model
subdirectory from `checkpoints.json`, rather than loading the repository root.

| Family | Models | Final step |
|---|---:|---:|
| GSM8K / Qwen2.5-1.5B-Instruct | 21 | 500 |
| Math / Qwen3-8B-SFT496 | 6 | 300 |
| Math / Qwen2.5-Math-7B | 6 | 500 |
| Coding / DeepSeek-Coder-1.3B-Base | 6 | 500 |

The manifest lists planned models; a model is available after its folder has
been committed. Historical hybrid runs retain their original configuration
name. Base-model licenses remain applicable; no optimizer states or training
datasets are published here.

```python
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer
subdir = "models/gsm8k/fixed2/seed42/step500"
root = snapshot_download("Gsj49/adaptiveG", allow_patterns=[subdir + "/*"])
folder = Path(root) / subdir
tokenizer = AutoTokenizer.from_pretrained(folder)
model = AutoModelForCausalLM.from_pretrained(folder)
```