💻 [Github Repo] • 📃 [Paper] • 🤗 [VaLProbing-32K]
**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).
