README.md

---
license: apache-2.0
language:
- en
---

# FILM-7B

<p align="center">
   💻 <a href="https://github.com/microsoft/FILM/" target="_blank">[Github Repo]</a> • 📃 <a href="https://arxiv.org/abs/2404.16811" target="_blank">[Paper]</a> • ⚓ <a href="https://huggingface.co/datasets/In2Training/VaLProbing-32K" target="_blank">[VaLProbing-32K] </a>
</p>

**FILM-7B is a 32K-context LLM that overcomes the lost-in-the-middle problem.**
It is trained from Mistral-7B-Instruct-v0.2 by applying Information-Intensie (In2) Training.
FILM-7B achieves near-perfect performance on probing tasks, SOTA-level performance on real-world long-context tasks among ~7B size LLMs, and does not compromise the short-context performance.

## Model Usage

The system tempelate for FILM-7B:
```text
'''[INST] Below is a context and an instruction. Based on the information provided in the context, write a response for the instruction.

### Context:
{YOUR LONG CONTEXT}

### Instruction:
{YOUR QUESTION & INSTRUCTION} [/INST]
'''
```

## Probing Results

To reproduce the results on our VaL Probing, see the guidance in [https://github.com/microsoft/FILM/tree/main/VaLProbing](https://github.com/microsoft/FILM/tree/main/VaLProbing).

<p align="center">
    <img src="./figures/probing_results_new.png" width="800">
    <br>
</p>

## Real-World Long-Context Tasks

To reproduce the results on real-world long-context tasks, see the guidance in [https://github.com/microsoft/FILM/tree/main/real_world_long](https://github.com/microsoft/FILM/tree/main/real_world_long).

<p align="center">
    <img src="./figures/real_world_long.png" width="800">
    <br>
</p>

## Short-Context Tasks

To reproduce the results on short-context tasks, see the guidance in [https://github.com/microsoft/FILM/tree/main/short_tasks](https://github.com/microsoft/FILM/tree/main/short_tasks).

<p align="center">
    <img src="./figures/short.png" width="800">
    <br>
</p>

## 📝 Citation
```
@misc{an2024make,
      title={Make Your LLM Fully Utilize the Context}, 
      author={Shengnan An and Zexiong Ma and Zeqi Lin and Nanning Zheng and Jian-Guang Lou},
      year={2024},
      eprint={2404.16811},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

Disclaimer: This model is strictly for research purposes, and not an official product or service from Microsoft.