Qwen3-Reranker-HomeDepot

Fine-tuned 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

pip install transformers torch

Basic Usage

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

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

@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
• Dataset: Home Depot Product Search Relevance
• Training framework: HuggingFace Transformers