---
base_model: meta-llama/Llama-3.2-3B-Instruct
license: llama3.2
tags:
  - sft
  - math
  - thinking
  - llama
---

# Llama-3.2-3B-ThinkSFT

`Llama-3.2-3B-Instruct` fine-tuned on 43.5K explicit reasoning traces
from [OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k)
(math subset, thinking format). No continual pre-training.

**Pipeline:** Base → Thinking SFT

Released as part of:
**[When Can LLMs Learn to Reason with Weak Supervision?](https://salmanrahman.net/rlvr-weak-supervision)**
— Rahman, Shen, Mordvina, Palangi, Gabriel, Izmailov (2026)

## Training Details

| | |
|---|---|
| Init | `meta-llama/Llama-3.2-3B-Instruct` |
| Data | [OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k) math subset (43.5K examples) |
| Epochs | 3 |
| Sequence length | 8,192 |
| Effective batch size | 256 sequences |
| Learning rate | 1.5e-5, cosine decay, 10% warmup |
| Optimizer | AdamW, weight decay 0.01 |
| Precision | BF16 + Flash Attention 2 |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("pavelslab-nyu/Llama-3.2-3B-ThinkSFT")
tokenizer = AutoTokenizer.from_pretrained("pavelslab-nyu/Llama-3.2-3B-ThinkSFT")
```

## Citation

```bibtex
@article{rahman2026when,
  title   = {When Can LLMs Learn to Reason with Weak Supervision?},
  author  = {Rahman, Salman and Shen, Jingyan and Mordvina, Anna and
             Palangi, Hamid and Gabriel, Saadia and Izmailov, Pavel},
  journal = {Preprint},
  year    = {2026}
}
```