README.md

---
license: apple-amlr
base_model:
- mistralai/Mistral-7B-Instruct-v0.2
tags:
- rag
- compression
- retrieval
- generation
---

# CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning

<div align="center">
  <img src="clara_logo.jpg" width="300"/>
</div>

<div align="center">
  <a href="https://arxiv.org/abs/2511.18659"><img src="https://img.shields.io/badge/arXiv-2511.18659-b31b1b.svg" alt="arXiv"></a>
  <a href="https://arxiv.org/abs/2511.18659"><img src="https://img.shields.io/badge/Paper-PDF-red.svg" alt="Paper"></a>
  <a href="https://github.com/apple/ml-clara"><img src="https://img.shields.io/badge/GitHub-Code-blue.svg" alt="GitHub"></a>
</div>


# CLaRa-7B-Base (Compression-16 & 128)

The CLaRa-7B-Base model is our foundational unified RAG model with built-in semantic document compression (16× and 128x).  
It provides a base compressor + generator capable of producing answers directly from compressed document representations.

**Training recipe:** Trained using QA-guided semantic compression and paraphrase consistency objectives.  
**Benchmarks:** Strong baseline performance across multi-hop QA tasks under a 16× compression ratio.

---

## More details and usage examples:

Paper: [CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning](https://arxiv.org/abs/2511.18659)  
GitHub: https://github.com/apple/ml-clara

---

## Example Usage

```python
from transformers import AutoModel

unirag = AutoModel.from_pretrained(
    "/mnt/ceph_rbd/model/CLaRa-7B-Base/compression-16",
    trust_remote_code=True
).to("cuda")

documents = [
    [
        "Weldenia is a monotypic genus of flowering plant in the family Commelinaceae...",
        "Hagsatera is a genus of orchids native to Mexico and Guatemala...",
        "Alsobia is a genus of flowering plants native to Mexico and Central America..."
    ]
]

questions = [""]

out = unirag.generate_from_paraphrase(
    questions=questions,
    documents=documents,
    max_new_tokens=64
)

print("Generated answer:", out)