---
base_model: unsloth/llama-3.3-70b-instruct-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- trl
license: apache-2.0
language:
- en
datasets:
- codelion/Sky-T1_data_17k
- codelion/optillm-router-dataset
metrics:
- accuracy
---

# Llama-3.3-70B-o1 Thinker Model

This model was fine-tuned on CoT-style reasoning traces. The model will respond with a _thinking_ trace
  between the `<|begin_of_thought|>` and `<|end_of_thought|>` tags. The final answer will be between the `<|begin_of_solution|>` 
and `<|end_of_solution|>` tags.

Compared to the base Llama model, this thinker model has a tendency to generate a large number of tokens. So, if you are benchmarking
make sure you have the full generated text in the reponse, ending with the `<|end_of_solution|>` tag. For most queries,
you will need to set the `max_tokens` to at least 8192.

The GGUF quants for the model are available here - [Llama-3.3-70B-o1-gguf](https://huggingface.co./codelion/Llama-3.3-70B-o1-gguf)

The model was trained using QLoRA fine-tuning. You can find the adapter here - [Llama-3.3-70B-o1-lora](https://huggingface.co./codelion/Llama-3.3-70B-o1-lora).

## Evaluation results

| Model | AIME 2024 pass@1 |
|-------|------------------|
| **Llama-3.3-70B-o1** | **46.7** |
| Llama-3.3-70B | 30.0 | 
| Sky-T1-32B-Preview | 43.3 |
| o1-preview | 40.0 |
| QwQ | 50.0 |

- **Developed by:** codelion
- **License:** apache-2.0
- **Finetuned from model :** unsloth/llama-3.3-70b-instruct-bnb-4bit

This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)