This model card describes a LoRA model presented in How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?.

The performance of Large Language Models (LLMs) on many tasks is greatly limited by the knowledge learned during pre-training and stored in the model's parameters. Low-rank adaptation (LoRA) is a popular and efficient training technique for updating or domain-specific adaptation of LLMs. In this study, we investigate how new facts can be incorporated into the LLM using LoRA without compromising the previously learned knowledge. We fine-tuned Llama-3.1-8B-instruct using LoRA with varying amounts of new knowledge. Our experiments have shown that the best results are obtained when the training data contains a mixture of known and new facts. However, this approach is still potentially harmful because the model's performance on external question-answering benchmarks declines after such fine-tuning. When the training data is biased towards certain entities, the model tends to regress to few overrepresented answers. In addition, we found that the model becomes more confident and refuses to provide an answer in only few cases. These findings highlight the potential pitfalls of LoRA-based LLM updates and underscore the importance of training data composition and tuning parameters to balance new knowledge integration and general model capabilities.

The model can be used for question answering.

The model can be further fine-tuned for domain-specific question answering.

The model may not perform well on questions outside the knowledge it has been fine-tuned on, or if the training data was biased.

The model may exhibit biases present in the training data. The model's performance may degrade on external question-answering benchmarks after fine-tuning, especially if the training data is biased towards certain entities.

Users should be aware of potential biases in the model's responses and the limitations of its knowledge.

[More Information Needed]

The training data consists of questions and answers generated using the head-to-tail pipeline with a Dbpedia script. See the paper and Github repository for more details. Model was trained on 3000 Unknown questions with 10 additional Paraphrased question per Unknown

The model was fine-tuned using LoRA.

LR = 1e-3
BS = 8
EPOCHS = 10
LoRA:
lora_rank = 1
lora_alpha = 2
use_rslora = True
lora_dropout = 0.1
bias = "none"
target_modules = ["down_proj", "gate_proj", "up_proj"]
task_type = "CAUSAL_LM"

For evaluation you can use notebooks from github repository

[More Information Needed]

@misc{pletenev2025knowledgepackloraadapter,
      title={How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?}, 
      author={Sergey Pletenev and Maria Marina and Daniil Moskovskiy and Vasily Konovalov and Pavel Braslavski and Alexander Panchenko and Mikhail Salnikov},
      year={2025},
      eprint={2502.14502},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.14502}, 
}

Pletenev, S., Marina, M., Moskovskiy, D., Konovalov, V., Braslavski, P., Panchenko, A., & Salnikov, M. (2025). How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?.