README.md

---
license: other
license_name: llama3.2
license_link: https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/LICENSE
base_model: unsloth/Llama-3.2-1B
tags:
  - math
  - fine-tuned
  - lora
  - unsloth
  - llama
datasets:
  - MathLLMs/MathCodeInstruct
language:
  - en
library_name: transformers
pipeline_tag: text-generation
model-index:
  - name: Llama-3.2-1B-MathCodeInstruct-{{SIZE}}
    results:
      - task:
          type: text-generation
          name: GSM8K
        dataset:
          type: gsm8k
          name: GSM8K
        metrics:
          - type: exact_match
            value: {{GSM8K_ACC}}
            name: exact match (flexible-extract, 5-shot)
      - task:
          type: text-generation
          name: ARC-Challenge
        dataset:
          type: ai2_arc
          name: ARC-Challenge
        metrics:
          - type: acc_norm
            value: {{ARC_ACC}}
            name: acc_norm (25-shot)
      - task:
          type: text-generation
          name: HellaSwag
        dataset:
          type: hellaswag
          name: HellaSwag
        metrics:
          - type: acc_norm
            value: {{HELLASWAG_ACC}}
            name: acc_norm (10-shot)
      - task:
          type: text-generation
          name: WinoGrande
        dataset:
          type: winogrande
          name: WinoGrande
        metrics:
          - type: acc
            value: {{WINOGRANDE_ACC}}
            name: acc (5-shot)
      - task:
          type: text-generation
          name: MMLU
        dataset:
          type: mmlu
          name: MMLU
        metrics:
          - type: acc
            value: {{MMLU_ACC}}
            name: acc (5-shot)
---

# Llama-3.2-1B-MathCodeInstruct-10k

A [Llama-3.2-1B](https://huggingface.co/unsloth/Llama-3.2-1B) fine-tune on **10k examples** from
[MathLLMs/MathCodeInstruct](https://huggingface.co/datasets/MathLLMs/MathCodeInstruct), trained to solve math
word problems with step-by-step natural-language reasoning interleaved with executable Python.

This is one of three sibling models trained on {5k, 10k, 20k}-example subsets of the same dataset, to study how
fine-tuning data volume trades off against both math performance and general capability. See the
[training write-up](https://github.com/OliverSundaram/finetuning-Llama3.2-1B) for the full comparison across all three.

## Training details

| |                                                                                         |
|---|-----------------------------------------------------------------------------------------|
| Base model | `unsloth/Llama-3.2-1B`                                                                  |
| Method | LoRA (r=16, α=16, dropout=0) on all attention + MLP projections, merged to full weights |
| Dataset | MathLLMs/MathCodeInstruct, 10k training examples                                        |
| Epochs | 1                                                                                       |
| Effective batch size | 16 (batch 1 × grad. accum. 16)                                                          |
| Learning rate | 2e-4, cosine schedule, warmup ratio 0.03                                                |
| Hardware | 1× RTX 4060 (8GB)                                                                       |
| Framework | Unsloth + TRL `SFTTrainer`                                                              |

## Benchmark results

All benchmarks run with [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness), each at
its standard published shot count, compared against the un-tuned base model.

| Benchmark | Llama-3.2-1B (base) | MathCodeInstruct-10k | Change |
|---|---|----------------------|---|
| GSM8K | 5.8% | 8.7%                 | 🟢 +2.9% |
| ARC-Challenge | 36.9% | 36.1%                | 🔴 -0.8% |
| HellaSwag | 64.2% | 63.8%                | 🔴 -0.4% |
| WinoGrande | 60.8% | 62.0%                | 🟢 +1.3% |

**Speed:** **40.53 tokens/sec**
(base model: 40.59 tokens/sec)

### MMLU by category

![MMLU comparison](assets/mmlu_10k.png)

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "OliverSundaram/Llama-3.2-1B-MathCodeInstruct-10k"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "Below is a math problem. Please solve it step by step."},
    {"role": "user", "content": "If a train travels 60 miles in 45 minutes, what is its speed in miles per hour?"},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## Limitations

- Trained on a single epoch of a 10k-example subset — not intended to be a general-purpose assistant.
- MMLU/ARC/HellaSwag/WinoGrande scores reflect a small 1B-parameter base model and should be read relative to
  the base model's own scores, not against much larger models.
- No safety alignment or RLHF was applied beyond what the base Llama-3.2-1B already has.