Qwen2.5-7B-query-rewriter-lora

Qwen 2.5 7B Query Rewriter - LoRA Adapter

Fine-tuned LoRA adapter for query rewriting in multi-turn conversational retrieval (MTRAGEval Task A).

Model Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Training Method: LoRA (Low-Rank Adaptation)
• Training Data: MTRAG Benchmark human evaluations (551 train, 62 validation)
• Best Checkpoint: Iteration 150
• Framework: MLX (Apple Silicon optimized)
• Task: Transform multi-turn conversational queries into standalone, search-friendly queries

Performance

The model resolves pronouns and includes necessary context from conversation history:

Usage

With MLX (Apple Silicon)

from mlx_lm import load, generate

# Load base model with adapter
model, tokenizer = load(
    "Qwen/Qwen2.5-7B-Instruct",
    adapter_path="caraman/Qwen2.5-7B-query-rewriter-lora"
)

# Prepare prompt
system_prompt = """You are a query rewriting assistant for information retrieval. Given a conversation history and a current question, rewrite the question to be completely standalone and self-contained."""

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": """CONVERSATION HISTORY:
USER: Tell me about the Eiffel Tower
ASSISTANT: The Eiffel Tower is in Paris, France.

CURRENT QUESTION: When was it built?

Rewrite this question to be standalone:"""}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
print(response)  # "When was the Eiffel Tower built?"

Training Configuration

• LoRA Rank: 16
• LoRA Alpha (scale): 32.0
• Dropout: 0.15
• Learning Rate: 1e-05
• Batch Size: 4 (effective 16 with gradient accumulation)
• Layers: 28
• Training Iterations: 200 (best at 150)

Training Data

Extracted from MTRAG Benchmark human query rewrites across 4 domains:

• ClapNQ (question answering)
• FiQA (finance)
• Government documents
• IBM Cloud documentation

Evaluation

Evaluated on 164 holdout conversation queries with nDCG@10 for retrieval performance.

Limitations

• Optimized for English only
• Best for technical/informational queries
• May not handle highly creative or open-ended questions well

Citation

@inproceedings{mtrageval2026,
  title={MTRAGEval: Multi-Turn Retrieval-Augmented Generation Evaluation},
  booktitle={SemEval 2026 Task 8},
  year={2026}
}

License

Apache 2.0