README.md

---
tags:
- reranker
- qwen3
- information-retrieval
- product-search
base_model: Qwen/Qwen3-Reranker-0.6B
license: mit
language:
- en
pipeline_tag: text-classification
---

# Qwen3-Reranker-HomeDepot

Fine-tuned [Qwen3-Reranker-0.6B](https://huggingface.co/Qwen/Qwen3-Reranker-0.6B) on the Home Depot product search dataset for e-commerce search ranking.

## Model Description

This model is a cross-encoder reranker trained to score query-product pairs for relevance. It takes a search query and product description as input and outputs a relevance score between 0 and 1.

**Base Model**: Qwen/Qwen3-Reranker-0.6B  
**Training Dataset**: Home Depot Product Search  
**Training Samples**: 51,911  
**Task**: Binary relevance classification (relevant/irrelevant)

## Training Details

### Dataset

- **Total samples**: 51,911
- **Splits**: 70% train / 15% validation / 15% test
- **Splitting strategy**: Query-stratified (prevents data leakage)
- **Label threshold**: Relevance ≥ 2.33 → relevant (1), else irrelevant (0)
- **Label distribution**: ~68% relevant, ~32% irrelevant

### Training Configuration

```
Learning rate: 5e-6
Batch size: 8 × 2 = 16
Epochs: 3
Optimizer: AdamW (weight_decay=0.01)
Scheduler: Linear warmup + decay
Mixed precision: BF16
```

### Hardware

- **GPU**: NVIDIA A100 80GB
- **Training time**: ~2-4 hours

## Usage

### Installation

```bash
pip install transformers torch
```

### Basic Usage

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

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "codefactory4791/Qwen3-Reranker-HomeDepot",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "codefactory4791/Qwen3-Reranker-HomeDepot",
    trust_remote_code=True
)

# Prepare input
query = "cordless drill"
document = "DEWALT 20V MAX Cordless Drill Kit with battery and charger"

# Format prompt
prompt = f'''<|im_start|>system
Judge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be "yes" or "no".<|im_end|>
<|im_start|>user
<Instruct>: Given a web search query, retrieve relevant passages that answer the query
<Query>: {query}
<Document>: {document}<|im_end|>
<|im_start|>assistant
<think>

</think>

'''

# Tokenize and get score
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model(**inputs)
logits = outputs.logits[0, -1, :]

# Get yes/no token probabilities
token_yes = tokenizer.convert_tokens_to_ids('yes')
token_no = tokenizer.convert_tokens_to_ids('no')
score = torch.sigmoid(logits[token_yes] - logits[token_no]).item()

print(f"Relevance score: {score:.4f}")
```

### Using with Ranking-Qwen Library

```python
from ranking_qwen.models import QwenReranker

# Load fine-tuned model
reranker = QwenReranker(model_name="codefactory4791/Qwen3-Reranker-HomeDepot")

# Score multiple candidates
scores = reranker.compute_scores(
    queries=["drill bits", "drill bits"],
    documents=[
        "DEWALT 14-Piece Titanium Drill Bit Set",
        "Black+Decker Screwdriver Set"
    ]
)
# Returns: [0.92, 0.31]
```

## Performance

Expected metrics on Home Depot test set:

- **NDCG@10**: ≥ 0.80
- **MAP**: ≥ 0.75
- **MRR**: ≥ 0.85
- **AUC**: ≥ 0.90

## Limitations

- Trained specifically for Home Depot product search
- May not generalize well to other domains without fine-tuning
- Maximum sequence length: 8192 tokens (though 2048 is recommended for speed)

## Citation

```bibtex
@misc{qwen3-reranker-homedepot,
  author = {Your Name},
  title = {Qwen3-Reranker Fine-tuned on Home Depot Dataset},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/codefactory4791/Qwen3-Reranker-HomeDepot}}
}
```

## License

MIT License - See base model license for additional details.

## Acknowledgments

- Base model: [Qwen3-Reranker-0.6B](https://huggingface.co/Qwen/Qwen3-Reranker-0.6B)
- Dataset: Home Depot Product Search Relevance
- Training framework: HuggingFace Transformers