Update README.md

README.md

@@ -1,57 +1,93 @@
 ---
 library_name: transformers
-license: mit
 pipeline_tag: question-answering
-tags:
-- lora
-- knowledge-editing
-- question-answering
+license: mit
+base_model: meta-llama/Llama-3.1-8B-Instruct
+tags: []
 ---
 
-# Model Card for Knowledge-Packed LoRA Adapters
+# Model Card for How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?
 
-This model card describes LoRA adapters fine-tuned to incorporate new knowledge into Large Language Models (LLMs), while preserving previously learned information. The approach and potential pitfalls of LoRA-based LLM updates are discussed in the paper: [How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?](https://arxiv.org/abs/2502.14502).
+This model card describes a LoRA model presented in [How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?](https://arxiv.org/abs/2502.14502).
 
 ## Model Details
 
-- **Developed by:** Sergey Pletenev, Maria Marina, Daniil Moskovskiy, Vasily Konovalov, Pavel Braslavski, Alexander Panchenko, and Mikhail Salnikov
-- **Model type:** LoRA adapter for causal language modeling
+### Model Description
+
+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.
+
+
+- **Developed by:** Sergey Pletenev, Maria Marina, Daniil Moskovskiy, Vasily Konovalov, Pavel Braslavski, Alexander Panchenko, Mikhail Salnikov
+- **Model type:** LLM
 - **Language(s) (NLP):** English
-- **License:** MIT
+- **License:** mit
 - **Finetuned from model:** meta-llama/Meta-Llama-3.1-8B-Instruct
 
+### Model Sources
+
+- **Repository:** [https://github.com/AIRI-Institute/knowledge-packing](https://github.com/AIRI-Institute/knowledge-packing)
+- **Paper:** [How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?](https://arxiv.org/abs/2502.14502)
+
+
 ## Uses
 
 ### Direct Use
-The model can be used to answer questions based on newly injected knowledge, for example, using facts from a specific domain. However, be mindful of the potential biases and knowledge spillover effects described in the paper.
+
+The model can be used for question answering.
+
+### Downstream Use
+
+The model can be further fine-tuned for domain-specific question answering.
 
 ### Out-of-Scope Use
-The model's performance may degrade when applied to tasks significantly different from the training data or when the training data is imbalanced. The model may exhibit biases learned from the training data and should not be used in high-stakes applications without careful evaluation and mitigation strategies.
+
+The model may not perform well on questions outside the knowledge it has been fine-tuned on, or if the training data was biased.
 
 ## Bias, Risks, and Limitations
-The model may regress to overrepresented answers when the training data is biased towards certain entities. Fine-tuning can negatively impact the model's performance on external question-answering benchmarks. The model may also become more confident and refuse to provide an answer in only a few cases.
 
+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.
 
-## How to Get Started with the Model
+### Recommendations
+
+Users should be aware of potential biases in the model's responses and the limitations of its knowledge.
 
-See the Github repository for instructions on generating the dataset and training LoRA adapters: [https://github.com/memyprokotow/lora_vs_persisted/tree/master](https://github.com/memyprokotow/lora_vs_persisted/tree/master)
+## How to Get Started with the Model
 
+[More Information Needed]
 
 ## Training Details
 
 ### Training Data
 
-The training data consists of a mixture of known and new facts, created using the head-to-tail pipeline with Dbpedia. The authors experimented with varying amounts of new knowledge. More details about the training data generation process can be found in the paper and the Github repo.  Datasets used for the paper can be downloaded from:
-- [Dataset with precollected triples and questions](https://drive.google.com/file/d/1pCtfRlvBW769384AgmfNBpIU8OmftfKd/view?usp=sharing)
-- [Questions with labelled knowledge categories](https://drive.google.com/file/d/1-NDeTa8TMRNY9UIsIqtI-Iw4vq-rda35/view?usp=sharing).
+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 500 Unknown questions with 10 additional HighlyKnown question per Unknown
 
 ### Training Procedure
 
-The model is fine-tuned using LoRA. 
+The model was fine-tuned using LoRA.
+
+#### Training Hyperparameters
+
+    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"
 
 ## Evaluation
 
-The model's performance was evaluated on external question-answering benchmarks and by analyzing knowledge spillover effects.  See the paper and Github repo for more details.
+For evaluation you can use [notebooks](https://github.com/AIRI-Institute/knowledge-packing/tree/main/notebooks) from github repository
+
+## Environmental Impact
+
+[More Information Needed]
+
 
 ## Citation
 
@@ -65,4 +101,8 @@ The model's performance was evaluated on external question-answering benchmarks
       primaryClass={cs.CL},
       url={https://arxiv.org/abs/2502.14502}, 
 }
-```
+```
+
+**APA:**
+
+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?.