In2Training_-_FILM-7B-gguf

Quantization made by Richard Erkhov.

Github

Discord

Request more models

FILM-7B - GGUF

• Model creator: https://huggingface.co/In2Training/
• Original model: https://huggingface.co/In2Training/FILM-7B/

Original model description:

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:

'''[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.

<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.

<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.

<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.