README.md

---
base_model: meta-llama/Llama-3.2-3B-Instruct
library_name: peft
license: other
tags:
- lora
- peft
- math
- reasoning
- adaption
- word-problems
- llama
- sft
datasets:
- Minutor/adaption-math-word-problem-sub-2
- Minutor/20k_math_dataset
pipeline_tag: text-generation
---

# adaption_math_word_problem_sub_2

## Model Training

A LORA adapter for `meta-llama/Llama-3.2-3B-Instruct`. This model was trained with SFT using [Adaption](https://adaptionlabs.ai)'s AutoScientist on the math_word_problem_sub_2 dataset.


![Training metrics](training-metrics.png)

### Model Details

- **Base model**: `meta-llama/Llama-3.2-3B-Instruct`
- **Training method**: SFT + LoRA
- **LoRA rank**: 16 | **alpha**: 32
- **Epochs**: 3
- **Learning rate**: 1e-5 (cosine schedule)
- **Trainable modules**: all-linear
- **Data format**: chat

### Training Data
The model was trained on 19,573 rows of adapted data with the following domain distribution: math (99%), language (0%), science (0%), personal-finance (0%), fitness-sports (0%), animal-nature (0%), agriculture (0%), how-to (0%), sports (0%), travel (0%), data-analysis-visualization (0%).

Trained on the adapted dataset:  
→ [Minutor/adaption-math-word-problem-sub-2](https://huggingface.co/datasets/Minutor/adaption-math-word-problem-sub-2)

Which itself was derived from the cleaned seed:  
→ [Minutor/20k_math_dataset](https://huggingface.co/datasets/Minutor/20k_math_dataset)  
(GSM8K + NuminaMath-1.5 + OpenMathInstruct-1)

### Evaluation Results

Win rates are computed by Adaption using **Gemini 3.1 Pro** as the judge.
![Win rates](win-rates.png)
| Evaluation | Sample size | Base | Adapted | Change |
|------------|-------------|------|---------|--------|
| Win-rate on training distribution | 200 held-out datapoints | 42 | **58** | **+16** |
| Math Win-rate (Adaption held-out) | 100 unseen datapoints across Math tasks | 51 | 50 | –1 |

The model shows a clear +16 point improvement on its training distribution while remaining essentially neutral on Adaption’s broader Math evaluation set.

### How to use

notebook snippet: [collab shared notebook](https://colab.research.google.com/drive/1E5yBG_7vgviVKPE7qJpbTwJRKLs6YbV_?usp=sharing)
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Llama-3.2-3B-Instruct"
ADAPTER = "Minutor/adaption_math_word_problem_sub_2"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=dtype, device_map="auto")
model = PeftModel.from_pretrained(base, ADAPTER)
# Optional: merge for faster inference
# model = model.merge_and_unload()

tokenizer = AutoTokenizer.from_pretrained(BASE)

messages = [
    {"role": "user", "content": "A store sells apples for $2 each and oranges for $3 each. If a customer buys 4 apples and 3 oranges, how much do they pay in total?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```
### AutoScientist Config

```json
{
  "job_id": "dfdd2990-1ef7-41e4-829d-2924d668ab65",
  "training_experiment_id": "19aa2607-d5ad-43b7-90b2-5e7d3745b627",
  "original_model_name": "meta-llama/Llama-3.2-3B-Instruct",
  "trained_model_name": "adaption_math_word_problem_sub_2",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 16,
    "n_evals": 5,
    "n_epochs": 3,
    "batch_size": "max",
    "lora_alpha": 32,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.1,
    "weight_decay": 0,
    "learning_rate": 0.00001,
    "max_grad_norm": 2,
    "base_model_size": "3B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "all-linear"
  }
}
```

<!-- 
## Model Evaluation

The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.


![Win rates](win-rates.png)

| Domain | Win rate vs. base model |
| --- | --- |
| math | 50% | -->