---
base_model: Qwen/Qwen3-0.6B-Base
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3
- math
- sft
- qlora
- merged-lora
- fp16
---

# qwen3-0.6b-math-l45-qlora-merged-fp16-v2

Merged fp16 model from `Qwen/Qwen3-0.6B-Base` plus a QLoRA SFT adapter trained for MATH level 4-5 problem solving.

This is the corrected SFT v2 run using the train split for training. Final 500-question evaluation is intentionally separate.

## Training Summary

- Base model: `Qwen/Qwen3-0.6B-Base`
- Train file: `/kaggle/input/datasets/anurhalizah/math-he/math_level45_train.parquet`
- Test file: `/kaggle/input/datasets/anurhalizah/math-he/math_level45_test.parquet`
- Train rows used: `3994`
- Train subset size: `3994`
- Smoke loss eval rows: `200`
- Max sequence length: `2048`
- Epochs: `3`
- Learning rate: `0.0002`
- Precision: fp16
- Merge method: `manual_lora_cpu`
- Merged LoRA matrices: `196`

## Prompt Format

```text
### System:
{system_prompt}

### Problem:
{problem}

### Solution:

```

System prompt:

```text
You are a precise mathematical problem solver.
Follow this exact output contract:
1. Solve the problem step by step with concise reasoning.
2. Use valid LaTeX math notation for mathematical expressions.
3. Preserve LaTeX commands such as \frac{...}{...}, \sqrt{...}, x^{...}, subscripts, equations, and inequalities.
4. Put the final answer on its own last line exactly in this form: Final Answer: \boxed{...}
5. Do not use Markdown code fences.
6. Do not switch to a different answer format.
```

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo_id = "zannvznn/qwen3-0.6b-math-l45-qlora-merged-fp16-v2"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
```