Update README.md

README.md

@@ -13,6 +13,33 @@ tags:
 The main objective of this model is to enhance performance in tasks related to medical dialogue
 and question-answering.
 
+- **Developed by:** [https://writer.com/](https://writer.com/);
+- **Model type:** Causal decoder-only;
+- **Language(s) (NLP):** English;
+- **License:** Apache 2.0;
+- **Finetuned from model:** [Palmyra-20B](https://huggingface.co/Writer/palmyra-large).
+
+### Model Source
+
+[Palmyra-Med: Instruction-Based Fine-Tuning of LLMs Enhancing Medical Domain Performance](https://dev.writer.com/docs/palmyra-med-instruction-based-fine-tuning-of-llms-enhancing-medical-domain-performance)
+
+
+## Uses
+
+
+### Out-of-Scope Use
+
+Production use without adequate assessment of risks and mitigation; any use cases which may be considered irresponsible or harmful. 
+
+## Bias, Risks, and Limitations
+
+Palmyra-Med-20B is mostly trained on English data, and will not generalize appropriately to other languages. Furthermore, as it is trained on a large-scale corpora representative of the web, it will carry the stereotypes and biases commonly encountered online.
+
+### Recommendations
+
+We recommend users of Palmyra-Med-20B to develop guardrails and to take appropriate precautions for any production use.
+
+
 ## Usage 
 The model is compatible with the huggingface `AutoModelForCausalLM` and can be easily run on a single 40GB A100.
 
@@ -65,8 +92,61 @@ print(output)
 # The use of vaccines has led to a significant reduction in the incidence and severity of many diseases, including measles, mumps, rubella, and polio. 
 ```
 
+##  Dataset
+For the fine-tuning of our LLMs, we used a custom-curated medical dataset that combines data from
+two publicly available sources: PubMedQA (Jin et al. 2019) and MedQA (Zhang et al. 2018).The
+PubMedQA dataset, which originated from the PubMed abstract database, consists of biomedical
+articles accompanied by corresponding question-answer pairs. In contrast, the MedQA dataset
+features medical questions and answers that are designed to assess the reasoning capabilities of
+medical question-answering systems.
+We prepared our custom dataset by merging and processing data from the aforementioned sources,
+maintaining the dataset mixture ratios detailed in Table 1. These ratios were consistent for finetuning
+both Palmyra-20b and Palmyra-40b models. Upon fine-tuning the models with this dataset, we refer
+to the resulting models as Palmyra-Med-20b and Palmyra-Med-40b, respectively.
+
+Dataset Ratio Count
+PubMedQA 75% 
+MedQA 25% 
+
+| Dataset | Ratio    | Count |
+| -----------|----------- | ----------- |
+| PubMedQA | 75%       | 150,000       |
+| MedQA | 25%    | 10,178        |
+
+
+## Evaluation
+we present the findings of our experiments, beginning with the evaluation outcomes of
+the fine-tuned models and followed by a discussion of the base models’ performance on each of the
+evaluation datasets. Additionally, we report the progressive improvement of the Palmyra-Med-40b
+model throughout the training process on the PubMedQA dataset.
+
+| Model | PubMedQA    | MedQA |
+| -----------|----------- | ----------- |
+| Palmyra-20b | 49.8       | 31.2       |
+| Palmyra-40b | 64.8 | 43.1| 
+| Palmyra-Med-20b|  75.6 | 44.6| 
+| Palmyra-Med-40b|  81.1 | 72.4| 
+
+
+
 ## Limitation
 The model may not operate efficiently beyond the confines of the healthcare field.
 Since it has not been subjected to practical scenarios, its real-time efficacy and precision remain undetermined.
 Under no circumstances should it replace the advice of a medical professional, and it must be regarded solely as a tool for research purposes.
 
+## Citation and Related Information
+
+
+To cite this model:
+```
+@misc{Palmyra-Med-20B,
+  author = {Writer Engineering team},
+  title = {{Palmyra-Large Parameter Autoregressive Language Model}},
+  howpublished = {\url{https://dev.writer.com}},
+  year = 2023,
+  month = March 
+}
+```
+
+## Contact
+Hello@writer.com