---
Llama3.1-Aloe-70B-Beta is an **open healthcare LLM** (released with a permissive CC-BY license) achieving **state-of-the-art performance** on several medical tasks. Aloe Beta is made available in two model sizes: [8B](https://huggingface.co/HPAI-BSC/Llama31-Aloe-Beta-8B) and [70B](https://huggingface.co/HPAI-BSC/Llama31-Aloe-Beta-70B). Both models are trained using the same recipe.
Aloe is trained on 20 medical tasks, resulting in a robust and versatile healthcare model. Evaluations show Aloe models to be among the best in their class. When combined with a RAG system ([also released](https://github.com/HPAI-BSC/prompt_engine)) the 8B version gets close to the performance of closed models like MedPalm-2, GPT4. With the same RAG system, Aloe-Beta-70B outperforms those private alternatives, producing state-of-the-art results.
# Aloe-70B-Beta
**Aloe-70B-Beta** is the latest iteration in the **Aloe family**, building and improving on the success of its predecessor, [Aloe-8B-Alpha](https://huggingface.co/HPAI-BSC/Llama3-Aloe-8B-Alpha) in a larger model size.
Beta more than **triples** the training data used by Alpha, for a total of **1.8B tokens**, including a wider variety of medical tasks and instructions (e.g., text summarization, explanation, diagnosis, text classification, treatment recommendation, ...).
To mitigate catastrophic forgetting and enable the model to effectively learn new capabilities like **function calling**, we incorporated a diverse set of high-quality general-purpose data constituting 20% of the total training set. The curated data includes some of the highest-quality content available across a range of topics, including mathematics, programming, STEM, and very long instructions (> 8k tokens), to enrich the model's adaptability and comprehension across diverse domains.
Beta also boosts the alignment and safety stages with respect to Alpha. This includes a [medical preference dataset](https://huggingface.co/datasets/TsinghuaC3I/UltraMedical-Preference), as well as the red-teaming dataset (available soon).
Complete training details, model merging configurations, and all training data (including synthetically generated data) can be found below. This includes [the RAG system](https://github.com/HPAI-BSC/prompt_engine) that was developed to test Aloe Beta in a deployment setup. Aloe comes with a healthcare-specific risk assessment to facilitate to the safe use and deployment of such systems.
## Model Details
### [](https://huggingface.co/templates/model-card-example#model-description)Model Description
- **Developed by:** [HPAI](https://hpai.bsc.es/)
- **Model type:** Causal decoder-only transformer language model
- **Language(s) (NLP):** English (capable but not formally evaluated on other languages)
- **License:** This model is based on Meta Llama 3.1 70B and is governed by the [Meta Llama 3 License](https://www.llama.com/llama3_1/license/). All our modifications are available with a [CC BY 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license, making the Aloe Beta models **compatible with commercial use**.
- **Base model :** [meta-llama/Llama-3.1-70B](https://huggingface.co/meta-llama/Llama-3.1-70B)
- **Paper:** (more coming soon)
- **RAG Repository:** https://github.com/HPAI-BSC/prompt_engine
### [](https://huggingface.co/templates/model-card-example#model-sources-optional)Model Sources [optional]
## Model Performance
Aloe Beta has been tested on the most popular healthcare QA datasets, with and without **Medprompt** inference technique. Results show competitive performance, achieving SOTA within models of the same size.
The Beta model has been developed to excel in several different medical tasks. For this reason, we evaluated the model in many different medical benchmarks:
We also compared the performance of the model in the general domain, using the OpenLLM Leaderboard benchmark. Aloe-Beta gets competitive results with the current SOTA general models in the most used general benchmarks and outperforms the medical models:
## Uses
### Direct Use
We encourage the use of Aloe for research purposes, as a stepping stone to build better foundational models for healthcare. In production, Aloe should always be used under the supervision of a human expert.
### Out-of-Scope Use
These models are not to be used for clinical practice, medical diagnosis, or any other form of direct or indirect healthcare advice. Models are prone to error and can produce toxic content. The use of Aloe models for activities harmful to individuals, such as spam, fraud, or impersonation, is strictly prohibited. Minors should not be left alone to interact with Aloe without supervision.
## Bias, Risks, and Limitations
Aloe can produce toxic content under the appropriate prompts, and it includes multiple undesirable biases. While significant efforts where conducted to mitigate this (see Alignment details below), model safety cannot be fully guaranteed. We avoid the use of all personal data in our training.
We identify at least three risk cases specific of healthcare LLMs:
- Healthcare professional impersonation, a fraudulent behaviour which currently generates billions of dollars in [profit](https://www.justice.gov/opa/pr/justice-department-charges-dozens-12-billion-health-care-fraud). A model such as Aloe could be used to increase the efficacy of such deceiving activities, making them more widespread. The main preventive actions are public literacy on the unreliability of digitised information and the importance of medical registration, and legislation enforcing AI-generated content disclaimers.
- Medical decision-making without professional supervision. While this is already an issue in modern societies (eg self-medication) a model such as Aloe, capable of producing high-quality conversational data, can facilitate self-delusion, particularly in the presence of sycophancy. By producing tailored responses, it can also be used to generate actionable answers. Public literacy on the dangers of self-diagnosis is one of the main defenses, together with the introduction of disclaimers and warnings on the models' outputs.
- Access to information on dangerous substances or procedures. While the literature on sensitive content can already be found on different sources (eg libraries, the internet, dark web), LLMs can centralize such access, making it nearly impossible to control the flow of such information. Model alignment can help in that regard, but so far the effects remain insufficient, as jailbreaking methods still overcome it.
## How to Get Started with the Model
Use the code below to get started with the model. You can run conversational inference using the Transformers pipeline abstraction, or by leveraging the Auto classes with the `generate()` function. Let's see examples for both.
#### Transformers pipeline
```python
import transformers
import torch
model_id = "HPAI-BSC/Llama3.1-Aloe-Beta-70B"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
messages = [
{"role": "system", "content": "You are an expert medical assistant named Aloe, developed by the High Performance Artificial Intelligence Group at Barcelona Supercomputing Center(BSC). You are to be a helpful, respectful, and honest assistant."},
{"role": "user", "content": "Hello."},
]
prompt = pipeline.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
terminators = [
pipeline.tokenizer.eos_token_id,
pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = pipeline(
prompt,
max_new_tokens=256,
eos_token_id=terminators,
do_sample=True,
temperature=0.6,
top_p=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])
```
#### Transformers AutoModelForCausalLM
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "HPAI-BSC/Llama3.1-Aloe-Beta-70B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [
{"role": "system", "content": "You are an expert medical assistant named Aloe, developed by the High Performance Artificial Intelligence Group at Barcelona Supercomputing Center(BSC). You are to be a helpful, respectful, and honest assistant."},
{"role": "user", "content": "Hello"},
]
input_ids = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
terminators = [
tokenizer.eos_token_id,
tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = model.generate(
input_ids,
max_new_tokens=256,
eos_token_id=terminators,
do_sample=True,
temperature=0.6,
top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
```
## Training Details
### Supervised fine-tuning
SFT on top of Llama 3.1 using axolotl (https://github.com/axolotl-ai-cloud/axolotl).
We used Deepspeed's Zero-3 distributed training using the following hardware:
* 8B: 32x NVIDIA Hopper H100 64GB of the *Marenostrum 5*.
* 70B: 64x NVIDIA Hopper H100 64GB of the *Marenostrum 5*.
#### Training Data
The training set consists of around 1.8B tokens, having 3 different types of data:
- Medical domain datasets. Includes data from 20 different medical tasks.
- [HPAI-BSC/Aloe-Beta-General-Collection](https://huggingface.co/datasets/HPAI-BSC/Aloe-Beta-General-Collection)
- [HPAI-BSC/chain-of-diagnosis](https://huggingface.co/datasets/HPAI-BSC/chain-of-diagnosis)
- [HPAI-BSC/MedS-Ins](https://huggingface.co/datasets/HPAI-BSC/MedS-Ins)
- [HPAI-BSC/ultramedica](https://huggingface.co/datasets/HPAI-BSC/ultramedical)
- Synthetic data. We expanded our training data by generating high-quality answers using Llama3.1-70B:
- [HPAI-BSC/pubmedqa-cot-llama31](https://huggingface.co/datasets/HPAI-BSC/pubmedqa-cot-llama31)
- [HPAI-BSC/medqa-cot-llama31](https://huggingface.co/datasets/HPAI-BSC/medqa-cot-llama31)
- [HPAI-BSC/medmcqa-cot-llama31](https://huggingface.co/datasets/HPAI-BSC/medmcqa-cot-llama31)
- [HPAI-BSC/headqa-cot-llama31](https://huggingface.co/datasets/HPAI-BSC/headqa-cot-llama31)
- [HPAI-BSC/MMLU-medical-cot-llama31](https://huggingface.co/datasets/HPAI-BSC/MMLU-medical-cot-llama31)
- [HPAI-BSC/Polymed-QA](https://huggingface.co/datasets/HPAI-BSC/Polymed-QA)
- General data. It includes maths, STEM, code, function calling, and instruction with very long context.
- [HPAI-BSC/Aloe-Beta-General-Collection](https://huggingface.co/datasets/HPAI-BSC/Aloe-Beta-General-Collection)
#### Training parameters
- Epochs: 4
- Sequence length: 16384
- Optimizer: adamw_torch
- Learning rate: 2e-5
- Learning rate scheduler: cosine
- Warmup steps: 100
- Weight decay: 0
- Gradient checkpointing
- Zero 3
- Total batch size: 128
- Batch size per device: 1
- Gradient accumulation steps: 2
### Model Merging
The model trained was merged with the Llama-3.1-Instruct model using the DARE_TIES technique. [Mergekit](https://github.com/arcee-ai/mergekit) was used to conduct the merging.
### Model Alignment
The model is aligned using the Direct Preference Optimization (DPO) technique through a two-step process:
1. General DPO Alignment: This step uses a dataset combining medical, general preference, and safety data. We used our dataset [HPAI-BSC/Aloe-Beta-DPO](https://huggingface.co/datasets/HPAI-BSC/Aloe-Beta-DPO). We split the dataset into five parts, and the model was trained iteratively for one epoch on each chunk. We used a learning rate of 2e-7.
2. Red-Teaming Alignment: This step further fine-tunes the model to resist a variety of potential attacks, enhancing its robustness and security. The dataset will be shared soon. In this stage, we set the learning rate to 1e-7.
We used [OpenRLHF](https://github.com/OpenRLHF/OpenRLHF) library. We aligned the model using 25x NVIDA HOOPER H100 64GB of the *Marenostrum 5*. Common hyperparameters:
- Sequence length: 4096
- Optimizer: Fused adam
- Total batch size 100
- Batch size per device: 1
- Gradient accumulation steps: 4
- Beta: 0.1
## Evaluation
### Testing Data, Factors & Metrics
#### Testing Data
- [ACI-BENCH](https://github.com/wyim/aci-bench)
- [MTS-Dialog](https://github.com/abachaa/MTS-Dialog)
- [MedText](https://huggingface.co/datasets/BI55/MedText)
- [Medical Text classification](https://www.kaggle.com/datasets/chaitanyakck/medical-text/data)
- [OLAPH](https://github.com/dmis-lab/OLAPH)
- CareQA Open
- [MedDialog](https://huggingface.co/datasets/bigbio/meddialog)
- [MEDIQA QA](https://huggingface.co/datasets/bigbio/mediqa_qa)
- [Meddialog Qsumm](https://huggingface.co/datasets/lighteval/med_dialog)
- [Biored](https://huggingface.co/datasets/YufeiHFUT/BioRED_all_info)
- [MIMIC-III](https://huggingface.co/datasets/dmacres/mimiciii-hospitalcourse-meta)
- [Medical Prescription](https://huggingface.co/datasets/devlocalhost/prescription-full)
- [MedQA (USMLE)](https://huggingface.co/datasets/bigbio/med_qa)
- [MedMCQA](https://huggingface.co/datasets/medmcqa)
- [PubMedQA](https://huggingface.co/datasets/bigbio/pubmed_qa)
- [MMLU-Medical](https://huggingface.co/datasets/lukaemon/mmlu)
- [MedQA-4-Option](https://huggingface.co/datasets/GBaker/MedQA-USMLE-4-options)
- [CareQA](https://huggingface.co/datasets/HPAI-BSC/CareQA)
- [Open LLM Leaderboard 2](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
#### Metrics
- Accuracy: suite the evaluation of multiple-choice question-answering tasks.
- Rouge1: refers to the overlap of unigrams between the system and the gold standard.
#### Summary
To compare Aloe with the most competitive open models (both general purpose and healthcare-specific) we use popular healthcare datasets (PubMedQA, MedMCQA, MedQA and MMLU for six medical tasks only), together with the new and highly reliable CareQA. However, while MCQA benchmarks provide valuable insights into a model's ability to handle structured queries, they fall short of representing the full range of challenges faced in medical practice. Building upon this idea, Aloe-Beta represents the next step in the evolution of the Aloe Family, designed to broaden the scope beyond the multiple-choice question-answering tasks that define Aloe-Alpha.
Benchmark results indicate the training conducted on Aloe has boosted its performance achieving comparable results with SOTA models like Llama3-OpenBioLLLM, Llama3-Med42, MedPalm-2 and GPT-4. Llama3.1-Aloe-Beta-70B also outperforms the other existing medical models in the OpenLLM Leaderboard and in the evaluation of other medical tasks like Medical Factualy and Medical Treatment recommendations among others. All these results make Llama3.1-Aloe-Beta-70B one of the best existing models for healthcare.
With the help of prompting techniques the performance of Llama3.1-Aloe-Beta-70B is significantly improved. Medprompting in particular provides a 4% increase in reported accuracy, after which Llama3.1-Aloe-Beta-70B outperforms all the existing models that do not use RAG evaluation.
## Environmental Impact
- **Hardware Type:** 64xH100
- **Hours used (8B):** 544 GPU hours
- **Hours used (70B):** 4500 GPU hours
- **Hardware Provider:** Barcelona Supercomputing Center (BSC)
- **Compute Region:** Spain
- **Carbon Emitted:** 34.1 kg of CO2
## Authors
Aloe Beta has been developed by the [High Performance Artificial Intelligence](https://hpai.bsc.es/) research group, from the [Barcelona Supercomping Center - BSC](https://www.bsc.es/). Main authors are [Jordi Bayarri Planas](https://huggingface.co/JordiBayarri), [Ashwin Kumar Gururajan](https://huggingface.co/G-AshwinKumar) and [Dario Garcia-Gasulla](https://huggingface.co/dariog). Red teaming efforts lead by Adrian Tormos.
mailto:hpai@bsc.es
## Citations
If you use this repository in a published work, please cite the corresponding papers as source:
```
@misc{gururajan2024aloe,
title={Aloe: A Family of Fine-tuned Open Healthcare LLMs},
author={Ashwin Kumar Gururajan and Enrique Lopez-Cuena and Jordi Bayarri-Planas and Adrian Tormos and Daniel Hinjos and Pablo Bernabeu-Perez and Anna Arias-Duart and Pablo Agustin Martin-Torres and Lucia Urcelay-Ganzabal and Marta Gonzalez-Mallo and Sergio Alvarez-Napagao and Eduard Ayguadé-Parra and Ulises Cortés Dario Garcia-Gasulla},
year={2024},
eprint={2405.01886},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```
