Model Description

This model is a fine-tuned version of Gemma 4 trained for retrieval and reranking tasks. Given a search query and a collection of candidate passages, the model selects and returns the most relevant passage from the provided corpus.

The model was fine-tuned using Unsloth and Hugging Face's TRL library for efficient training.

Input Format

The model was trained on chat-formatted examples:

python
Copy code
[
    {"role": "system", "content": "<|GET|>"},
    {
        "role": "user",
        "content": {
            "query": "most dependable affordable cars",
            "corpus": [
                "Document 1...",
                "Document 2...",
                "If you can look past its bargain interior and anonymous exterior, the Suzuki SX4 is one of the most reliable and affordable all-wheel-drive cars."
            ]
        }
    }
]

Output Format

The model returns the most relevant passage from the corpus:

text
Copy code
If you can look past its bargain interior and anonymous exterior, the Suzuki SX4 is one of the most reliable and affordable all-wheel-drive cars.

Training Objective

• Query-document relevance matching
• Passage retrieval and reranking
• Selection of the best matching document from a candidate set

Training Framework

• Unsloth
• Hugging Face Transformers
• Hugging Face TRL

Acknowledgements

This model was fine-tuned using Unsloth for fast and memory-efficient training.