---
language:
- en
license: other
license_name: qwen-research
base_model: Qwen/Qwen2.5-8B
tags:
- qwen
- lora
- merged
- math
- reasoning
- tool-integrated-reasoning
- aimo
- safetensors
datasets:
- jeannkouagou/aimo3-tool-integrated-reasoning
pipeline_tag: text-generation
library_name: transformers
---

# Qwen3-8B AIMO3 Tool-Integrated Reasoning

## Model Summary

A LoRA fine-tuned version of Qwen-8B trained for **tool-integrated reasoning** on the AIMO3 competition dataset (generated by GPT-OSS-120B). The LoRA adapters have been **merged** into the base model and saved in SafeTensors format for straightforward deployment.

| Property | Details |
|---|---|
| Base Model | Qwen-8B |
| Fine-tuning Method | LoRA (merged) |
| Format | SafeTensors (BF16) |
| Parameters | ~8B |
| Disk Size | ~16GB |
| Max Context | 8192 tokens |

---

## Model Details

### LoRA Configuration

| Hyperparameter | Value |
|---|---|
| Rank (r) | 16 |
| Alpha | 32 |
| Dropout | 0.05 |
| Bias | none |
| Target Modules | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` |

### Training Hyperparameters

| Hyperparameter | Value |
|---|---|
| Precision | BFloat16 (no quantization) |
| Epochs | 2 |
| Steps | 8750 (~1 epoch) |
| Per-device Batch Size | 2 |
| Gradient Accumulation Steps | 8 (effective batch: 16) |
| Learning Rate | 2e-4 |
| LR Scheduler | Cosine with warmup |
| Warmup Ratio | 0.03 |
| Weight Decay | 0.01 |
| Max Gradient Norm | 1.0 |
| Max Sequence Length | 8192 |
| Optimizer | AdamW (Fused) |

### Hardware & Infrastructure

- **Platform**: Kaggle
- **GPU**: Single NVIDIA H100 (80GB)
- **Attention**: Flash Attention 2
- **Optimizations**: Gradient checkpointing, TF32, fused optimizer

---

## Training Data

- **Dataset**: [AIMO3 Tool-Integrated Reasoning Dataset](https://www.kaggle.com/datasets/jeannkouagou/aimo3-tool-integrated-reasoning) (synthesized by GPT-OSS-120B)
- **Split**: 97.5% train / 2.5% validation
- **Format**: CSV with problem–solution pairs

**Supported column names:**
- Input: `problem`, `question`, `input`, `prompt`
- Output: `solution`, `answer`, `output`, `response`, `completion`

### Instruction Format

Training uses a ChatML-style format:
```
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
{response}<|im_end|>
```

### Training Loss

The model is trained for 8750 steps (~ 1 epoch) before stopping. Below are the train and validation loss curves for the entire training session.

![Training Loss Plot](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F32392416%2Ffc9c0a306042ea3b976ed8749150a48d%2Floss_plot.png?generation=1774706655261800&alt=media)

---

## Usage

### Load the Model

Since the LoRA adapters are already merged, PEFT is **not required**:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "tensorhydra/qwen-8b-aimo3-reasoning",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained(
    "tensorhydra/qwen-8b-aimo3-reasoning",
    trust_remote_code=True
)
```

### Inference
```python
prompt = "Solve this problem: What is 2 + 2?"

formatted_prompt = f"user\n{prompt}\nassistant\n"

inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=False)
print(response)
```

### Batch Inference
```python
prompts = [
    "Solve: 15 + 27 = ?",
    "What is the derivative of x^2?",
    "Calculate the area of a circle with radius 5"
]

formatted_prompts = [
    f"user\n{p}\nassistant\n"
    for p in prompts
]

inputs = tokenizer(formatted_prompts, return_tensors="pt", padding=True).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)

for response in tokenizer.batch_decode(outputs, skip_special_tokens=False):
    print(response)
    print("-" * 80)
```

### Quantized Inference (Lower VRAM)
```python
# 8-bit (~8GB VRAM)
model = AutoModelForCausalLM.from_pretrained(
    "tensorhydra/qwen-8b-aimo3-reasoning",
    load_in_8bit=True,
    device_map="auto"
)

# 4-bit (~4GB VRAM)
model = AutoModelForCausalLM.from_pretrained(
    "tensorhydra/qwen-8b-aimo3-reasoning",
    load_in_4bit=True,
    device_map="auto"
)
```

---

## Memory Requirements

| Mode | VRAM |
|---|---|
| BF16 (full) | ~16GB |
| 8-bit quantized | ~8GB |
| 4-bit quantized | ~4GB |

---

## Repository Structure
```
model/
├── config.json
├── generation_config.json
├── model.safetensors.index.json
├── model-00001-of-0000X.safetensors
├── ...
├── tokenizer_config.json
├── tokenizer.json
└── special_tokens_map.json
```

---

## Intended Use

- Mathematical reasoning and problem solving  
- Tool-integrated step-by-step reasoning  
- Educational and research applications  
- Production deployment (merged model, no PEFT dependency)

## Limitations

- Fine-tuned on a narrow reasoning domain; may not generalize well to other tasks
- Hard context limit of 8192 tokens
- Performance is bounded by the quality and distribution of the synthetic training data
- Full merged model requires ~16GB storage (vs. ~100–200MB for LoRA adapters alone)

---

## Links

- **Dataset**: [jeannkouagou/aimo3-tool-integrated-reasoning](https://www.kaggle.com/datasets/jeannkouagou/aimo3-tool-integrated-reasoning)
- **Fine-tuning Notebook**: [tensorhydra/qwen3-8b-aimo3-finetune](https://www.kaggle.com/code/tensorhydra/qwen3-8b-aimo3-finetune)

---

## Citation
```bibtex
@misc{qwen-lora-aimo3,
  title   = {Qwen-8B LoRA Fine-tuned for Tool-Integrated Reasoning},
  author  = {tensorhydra},
  year    = {2025},
  howpublished = {Kaggle Model Hub},
  note    = {Merged LoRA model in SafeTensors format}
}
```

---

## Acknowledgements

- **Base model**: [Qwen-8B](https://huggingface.co/Qwen) by Alibaba Cloud
- **Training frameworks**: Hugging Face Transformers & PEFT
- **Dataset synthesis**: GPT-OSS-120B
- **Serialization**: SafeTensors
- **Training platform**: Kaggle (H100 GPU)

## License

This model inherits the license of the base Qwen-8B model. Please refer to the [Qwen license terms](https://huggingface.co/Qwen/Qwen2.5-8B/blob/main/LICENSE) before use.