---
license: llama3.2
datasets:
- a-m-team/AM-DeepSeek-R1-Distilled-1.4M
language:
- en
base_model:
- meta-llama/Llama-3.2-3B
---
# Llama 3.2 3B Reasoning Model 

## Model Details

**Base Model:** Meta Llama 3.2 3B
**Fine-tuning:** Full-weight training on 100k DeepSeek R1 reasoning examples
**Training Infrastructure:** MI300X with bf16 precision
**Context Length:** 131,072 tokens
**Reasoning Format:** Structured thinking with `<think></think>` tags

## Usage 

This repository contains the Q4_K_M GGUF version of the model, ready for use with Ollama or llama.cpp.

### Sampling Parameters
```bash
./llama-cli -m checkpoint-11500-Q4_K_M.gguf \
  --temp 0.3 \
  --top-p 0.9 \
  --top-k 40 \
  --repeat-penalty 1.15 \
  -p "Your prompt here" \
  -n 1024
```

## Expected Output Format

The model will structure its responses with reasoning tags:

```
<think>
Let me solve this step by step...
Speed = Distance / Time
Speed = 300km / 4 hours = 75 km/h
</think>
The average speed of the train is 75 km/h (kilometers per hour).
```

## Model Capabilities

**Strengths:**
- Mathematical reasoning and calculations
- Step-by-step problem solving
- Logical analysis and deduction
- Code reasoning and debugging
- Scientific problem solving

**Limitations:**
- May generate verbose reasoning for simple questions
- Occasional repetition in thinking process
- Not trained for specific domain knowledge beyond general reasoning

## License

This model is based on Llama 3.2 and follows Meta's licensing terms.