ethicalabs/xLSTM-7b-Polymath

Model Card for xlstm-7b-instruct-phase-2

[!WARNING] This repository contains experimental models designed strictly for academic evaluation and research purposes. Critical Constraints: No Production Deployment: Experimental models must not be deployed in commercial, enterprise, or mission-critical environments under any circumstances. No Liability: Experimental models are provided "as-is" without warranties of any kind. The developers assume zero liability for downstream consequences, system integration failures, or regulatory non-compliance resulting from unauthorized deployment.

This model is a fine-tuned version of ethicalabs/xLSTM-7b-Instruct for task alignment.

It has been trained using TRL using SFT on assistant-only tokens.

The k_proj and v_proj matrices have been frozen to isolate and preserve the model's pre-trained knowledge base.

This fine-tuning focused only on the q_proj (query) and FFN matrices, adapting the model's reasoning and query-retrieval mechanisms without overwriting its core, frozen knowledge.

This experiment was designed to test the hypothesis that the model's reasoning capabilities (q_proj) could be specialized for math/code while its knowledge (k_proj, v_proj) remained intact.

Quick start

Work in Progress!

Training procedure

<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>

This model was trained with SFT.

Evaluation

This model has been loaded in 4-bit and evaluated with lighteval

Framework versions

• —PEFT 0.17.1
• —TRL: 0.24.0
• —Transformers: 4.57.1
• —Pytorch: 2.8.0+cu126
• —Datasets: 4.2.0
• —Tokenizers: 0.22.1

Citations

Cite TRL as:

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}