ethicalabs/xLSTM-7b-Polymath

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

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

This model was trained with SFT.

Evaluation

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

Task	Version	Metric	Value		Stderr
all		acc	0.5383	±	0.1476
		acc:logprob_normalization=LogProbCharNorm(name='norm', ignore_first_space=True)	0.7000	±	0.1528
		acc:logprob_normalization=LogProbCharNorm(name='norm', ignore_first_space=False)	0.8000	±	0.1333
		truthfulqa_mc1	0.6000	±	0.1633
		truthfulqa_mc2	0.7066	±	0.1481
		em:normalize_gold=<function gsm8k_normalizer at 0x7c5d972c3ba0>&normalize_pred=<function gsm8k_normalizer at 0x7c5d972c3ba0>	0.6000	±	0.1633
leaderboard:arc:challenge:25		acc	0.8000	±	0.1333
		acc:logprob_normalization=LogProbCharNorm(name='norm', ignore_first_space=True)	0.7000	±	0.1528
leaderboard:gsm8k:5		em:normalize_gold=<function gsm8k_normalizer at 0x7c5d972c3ba0>&normalize_pred=<function gsm8k_normalizer at 0x7c5d972c3ba0>	0.6000	±	0.1633
leaderboard:hellaswag:10		acc	0.5000	±	0.1667
		acc:logprob_normalization=LogProbCharNorm(name='norm', ignore_first_space=False)	0.8000	±	0.1333
leaderboard:mmlu:_average:5		acc	0.5316	±	0.1474
leaderboard:mmlu:abstract_algebra:5		acc	0.3000	±	0.1528
leaderboard:mmlu:anatomy:5		acc	0.3000	±	0.1528
leaderboard:mmlu:astronomy:5		acc	0.7000	±	0.1528
leaderboard:mmlu:business_ethics:5		acc	0.4000	±	0.1633
leaderboard:mmlu:clinical_knowledge:5		acc	0.7000	±	0.1528
leaderboard:mmlu:college_biology:5		acc	0.5000	±	0.1667
leaderboard:mmlu:college_chemistry:5		acc	0.4000	±	0.1633
leaderboard:mmlu:college_computer_science:5		acc	0.4000	±	0.1633
leaderboard:mmlu:college_mathematics:5		acc	0.2000	±	0.1333
leaderboard:mmlu:college_medicine:5		acc	0.5000	±	0.1667
leaderboard:mmlu:college_physics:5		acc	0.5000	±	0.1667
leaderboard:mmlu:computer_security:5		acc	0.9000	±	0.1000
leaderboard:mmlu:conceptual_physics:5		acc	0.4000	±	0.1633
leaderboard:mmlu:econometrics:5		acc	0.4000	±	0.1633
leaderboard:mmlu:electrical_engineering:5		acc	0.7000	±	0.1528
leaderboard:mmlu:elementary_mathematics:5		acc	0.3000	±	0.1528
leaderboard:mmlu:formal_logic:5		acc	0.3000	±	0.1528
leaderboard:mmlu:global_facts:5		acc	0.3000	±	0.1528
leaderboard:mmlu:high_school_biology:5		acc	0.9000	±	0.1000
leaderboard:mmlu:high_school_chemistry:5		acc	0.5000	±	0.1667
leaderboard:mmlu:high_school_computer_science:5		acc	0.6000	±	0.1633
leaderboard:mmlu:high_school_european_history:5		acc	0.7000	±	0.1528
leaderboard:mmlu:high_school_geography:5		acc	1.0000	±	0.0000
leaderboard:mmlu:high_school_government_and_politics:5		acc	0.8000	±	0.1333
leaderboard:mmlu:high_school_macroeconomics:5		acc	0.6000	±	0.1633
leaderboard:mmlu:high_school_mathematics:5		acc	0.3000	±	0.1528
leaderboard:mmlu:high_school_microeconomics:5		acc	0.7000	±	0.1528
leaderboard:mmlu:high_school_physics:5		acc	0.3000	±	0.1528
leaderboard:mmlu:high_school_psychology:5		acc	0.9000	±	0.1000
leaderboard:mmlu:high_school_statistics:5		acc	0.5000	±	0.1667
leaderboard:mmlu:high_school_us_history:5		acc	0.8000	±	0.1333
leaderboard:mmlu:high_school_world_history:5		acc	0.9000	±	0.1000
leaderboard:mmlu:human_aging:5		acc	0.5000	±	0.1667
leaderboard:mmlu:human_sexuality:5		acc	0.4000	±	0.1633
leaderboard:mmlu:international_law:5		acc	0.6000	±	0.1633
leaderboard:mmlu:jurisprudence:5		acc	0.6000	±	0.1633
leaderboard:mmlu:logical_fallacies:5		acc	0.4000	±	0.1633
leaderboard:mmlu:machine_learning:5		acc	0.5000	±	0.1667
leaderboard:mmlu:management:5		acc	0.5000	±	0.1667
leaderboard:mmlu:marketing:5		acc	0.8000	±	0.1333
leaderboard:mmlu:medical_genetics:5		acc	0.9000	±	0.1000
leaderboard:mmlu:miscellaneous:5		acc	0.5000	±	0.1667
leaderboard:mmlu:moral_disputes:5		acc	0.7000	±	0.1528
leaderboard:mmlu:moral_scenarios:5		acc	0.1000	±	0.1000
leaderboard:mmlu:nutrition:5		acc	0.6000	±	0.1633
leaderboard:mmlu:philosophy:5		acc	0.5000	±	0.1667
leaderboard:mmlu:prehistory:5		acc	0.4000	±	0.1633
leaderboard:mmlu:professional_accounting:5		acc	0.3000	±	0.1528
leaderboard:mmlu:professional_law:5		acc	0.4000	±	0.1633
leaderboard:mmlu:professional_medicine:5		acc	0.2000	±	0.1333
leaderboard:mmlu:professional_psychology:5		acc	0.3000	±	0.1528
leaderboard:mmlu:public_relations:5		acc	0.3000	±	0.1528
leaderboard:mmlu:security_studies:5		acc	0.3000	±	0.1528
leaderboard:mmlu:sociology:5		acc	0.8000	±	0.1333
leaderboard:mmlu:us_foreign_policy:5		acc	0.7000	±	0.1528
leaderboard:mmlu:virology:5		acc	0.5000	±	0.1667
leaderboard:mmlu:world_religions:5		acc	0.8000	±	0.1333
leaderboard:truthfulqa:mc:0		truthfulqa_mc1	0.6000	±	0.1633
		truthfulqa_mc2	0.7066	±	0.1481
leaderboard:winogrande:5		acc	0.7000	±	0.1528

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}}
}