---
base_model: epfl-llm/meditron-70b
datasets:
- bigbio/med_qa
- medmcqa
- bigbio/pubmed_qa
- epfl-llm/guidelines
inference: false
language:
- en
license: llama2
metrics:
- accuracy
- perplexity
model_creator: EPFL LLM Team
model_name: Meditron 70B
model_type: llama
pipeline_tag: text-generation
prompt_template: '<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
'
quantized_by: TheBloke
tags:
- medical
- health
- llama2
---
# Meditron 70B - GGUF
- Model creator: [EPFL LLM Team](https://huggingface.co/epfl-llm)
- Original model: [Meditron 70B](https://huggingface.co/epfl-llm/meditron-70b)
## Description
This repo contains GGUF format model files for [EPFL LLM Team's Meditron 70B](https://huggingface.co/epfl-llm/meditron-70b).
These files were quantised using hardware kindly provided by [Massed Compute](https://massedcompute.com/).
### About GGUF
GGUF is a new format introduced by the llama.cpp team on August 21st 2023. It is a replacement for GGML, which is no longer supported by llama.cpp.
Here is an incomplete list of clients and libraries that are known to support GGUF:
* [llama.cpp](https://github.com/ggerganov/llama.cpp). The source project for GGUF. Offers a CLI and a server option.
* [text-generation-webui](https://github.com/oobabooga/text-generation-webui), the most widely used web UI, with many features and powerful extensions. Supports GPU acceleration.
* [KoboldCpp](https://github.com/LostRuins/koboldcpp), a fully featured web UI, with GPU accel across all platforms and GPU architectures. Especially good for story telling.
* [GPT4All](https://gpt4all.io/index.html), a free and open source local running GUI, supporting Windows, Linux and macOS with full GPU accel.
* [LM Studio](https://lmstudio.ai/), an easy-to-use and powerful local GUI for Windows and macOS (Silicon), with GPU acceleration. Linux available, in beta as of 27/11/2023.
* [LoLLMS Web UI](https://github.com/ParisNeo/lollms-webui), a great web UI with many interesting and unique features, including a full model library for easy model selection.
* [Faraday.dev](https://faraday.dev/), an attractive and easy to use character-based chat GUI for Windows and macOS (both Silicon and Intel), with GPU acceleration.
* [llama-cpp-python](https://github.com/abetlen/llama-cpp-python), a Python library with GPU accel, LangChain support, and OpenAI-compatible API server.
* [candle](https://github.com/huggingface/candle), a Rust ML framework with a focus on performance, including GPU support, and ease of use.
* [ctransformers](https://github.com/marella/ctransformers), a Python library with GPU accel, LangChain support, and OpenAI-compatible AI server. Note, as of time of writing (November 27th 2023), ctransformers has not been updated in a long time and does not support many recent models.
## Repositories available
* [AWQ model(s) for GPU inference.](https://huggingface.co/TheBloke/meditron-70B-AWQ)
* [GPTQ models for GPU inference, with multiple quantisation parameter options.](https://huggingface.co/TheBloke/meditron-70B-GPTQ)
* [2, 3, 4, 5, 6 and 8-bit GGUF models for CPU+GPU inference](https://huggingface.co/TheBloke/meditron-70B-GGUF)
* [EPFL LLM Team's original unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/epfl-llm/meditron-70b)
## Prompt template: ChatML
```
<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
```
## Compatibility
These quantised GGUFv2 files are compatible with llama.cpp from August 27th onwards, as of commit [d0cee0d](https://github.com/ggerganov/llama.cpp/commit/d0cee0d36d5be95a0d9088b674dbb27354107221)
They are also compatible with many third party UIs and libraries - please see the list at the top of this README.
## Explanation of quantisation methods
Click to see details
The new methods available are:
* GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
* GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
* GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
* GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
* GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
Refer to the Provided Files table below to see what files use which methods, and how.
## Provided files
| Name | Quant method | Bits | Size | Max RAM required | Use case |
| ---- | ---- | ---- | ---- | ---- | ----- |
| [meditron-70b.Q2_K.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q2_K.gguf) | Q2_K | 2 | 29.28 GB| 31.78 GB | smallest, significant quality loss - not recommended for most purposes |
| [meditron-70b.Q3_K_S.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q3_K_S.gguf) | Q3_K_S | 3 | 29.92 GB| 32.42 GB | very small, high quality loss |
| [meditron-70b.Q3_K_M.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q3_K_M.gguf) | Q3_K_M | 3 | 33.19 GB| 35.69 GB | very small, high quality loss |
| [meditron-70b.Q3_K_L.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q3_K_L.gguf) | Q3_K_L | 3 | 36.15 GB| 38.65 GB | small, substantial quality loss |
| [meditron-70b.Q4_0.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q4_0.gguf) | Q4_0 | 4 | 38.87 GB| 41.37 GB | legacy; small, very high quality loss - prefer using Q3_K_M |
| [meditron-70b.Q4_K_S.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q4_K_S.gguf) | Q4_K_S | 4 | 39.07 GB| 41.57 GB | small, greater quality loss |
| [meditron-70b.Q4_K_M.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q4_K_M.gguf) | Q4_K_M | 4 | 41.42 GB| 43.92 GB | medium, balanced quality - recommended |
| [meditron-70b.Q5_0.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q5_0.gguf) | Q5_0 | 5 | 47.46 GB| 49.96 GB | legacy; medium, balanced quality - prefer using Q4_K_M |
| [meditron-70b.Q5_K_S.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q5_K_S.gguf) | Q5_K_S | 5 | 47.46 GB| 49.96 GB | large, low quality loss - recommended |
| [meditron-70b.Q5_K_M.gguf](https://huggingface.co/TheBloke/meditron-70B-GGUF/blob/main/meditron-70b.Q5_K_M.gguf) | Q5_K_M | 5 | 48.75 GB| 51.25 GB | large, very low quality loss - recommended |
| meditron-70b.Q6_K.gguf | Q6_K | 6 | 56.59 GB| 59.09 GB | very large, extremely low quality loss |
| meditron-70b.Q8_0.gguf | Q8_0 | 8 | 73.29 GB| 75.79 GB | very large, extremely low quality loss - not recommended |
**Note**: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.
### Q6_K and Q8_0 files are split and require joining
**Note:** HF does not support uploading files larger than 50GB. Therefore I have uploaded the Q6_K and Q8_0 files as split files.
Click for instructions regarding Q6_K and Q8_0 files
### q6_K
Please download:
* `meditron-70b.Q6_K.gguf-split-a`
* `meditron-70b.Q6_K.gguf-split-b`
### q8_0
Please download:
* `meditron-70b.Q8_0.gguf-split-a`
* `meditron-70b.Q8_0.gguf-split-b`
To join the files, do the following:
Linux and macOS:
```
cat meditron-70b.Q6_K.gguf-split-* > meditron-70b.Q6_K.gguf && rm meditron-70b.Q6_K.gguf-split-*
cat meditron-70b.Q8_0.gguf-split-* > meditron-70b.Q8_0.gguf && rm meditron-70b.Q8_0.gguf-split-*
```
Windows command line:
```
COPY /B meditron-70b.Q6_K.gguf-split-a + meditron-70b.Q6_K.gguf-split-b meditron-70b.Q6_K.gguf
del meditron-70b.Q6_K.gguf-split-a meditron-70b.Q6_K.gguf-split-b
COPY /B meditron-70b.Q8_0.gguf-split-a + meditron-70b.Q8_0.gguf-split-b meditron-70b.Q8_0.gguf
del meditron-70b.Q8_0.gguf-split-a meditron-70b.Q8_0.gguf-split-b
```
## How to download GGUF files
**Note for manual downloaders:** You almost never want to clone the entire repo! Multiple different quantisation formats are provided, and most users only want to pick and download a single file.
The following clients/libraries will automatically download models for you, providing a list of available models to choose from:
* LM Studio
* LoLLMS Web UI
* Faraday.dev
### In `text-generation-webui`
Under Download Model, you can enter the model repo: TheBloke/meditron-70B-GGUF and below it, a specific filename to download, such as: meditron-70b.Q4_K_M.gguf.
Then click Download.
### On the command line, including multiple files at once
I recommend using the `huggingface-hub` Python library:
```shell
pip3 install huggingface-hub
```
Then you can download any individual model file to the current directory, at high speed, with a command like this:
```shell
huggingface-cli download TheBloke/meditron-70B-GGUF meditron-70b.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False
```
More advanced huggingface-cli download usage (click to read)
You can also download multiple files at once with a pattern:
```shell
huggingface-cli download TheBloke/meditron-70B-GGUF --local-dir . --local-dir-use-symlinks False --include='*Q4_K*gguf'
```
For more documentation on downloading with `huggingface-cli`, please see: [HF -> Hub Python Library -> Download files -> Download from the CLI](https://huggingface.co/docs/huggingface_hub/guides/download#download-from-the-cli).
To accelerate downloads on fast connections (1Gbit/s or higher), install `hf_transfer`:
```shell
pip3 install hf_transfer
```
And set environment variable `HF_HUB_ENABLE_HF_TRANSFER` to `1`:
```shell
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download TheBloke/meditron-70B-GGUF meditron-70b.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False
```
Windows Command Line users: You can set the environment variable by running `set HF_HUB_ENABLE_HF_TRANSFER=1` before the download command.
## Example `llama.cpp` command
Make sure you are using `llama.cpp` from commit [d0cee0d](https://github.com/ggerganov/llama.cpp/commit/d0cee0d36d5be95a0d9088b674dbb27354107221) or later.
```shell
./main -ngl 35 -m meditron-70b.Q4_K_M.gguf --color -c 4096 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant"
```
Change `-ngl 32` to the number of layers to offload to GPU. Remove it if you don't have GPU acceleration.
Change `-c 4096` to the desired sequence length. For extended sequence models - eg 8K, 16K, 32K - the necessary RoPE scaling parameters are read from the GGUF file and set by llama.cpp automatically. Note that longer sequence lengths require much more resources, so you may need to reduce this value.
If you want to have a chat-style conversation, replace the `-p ` argument with `-i -ins`
For other parameters and how to use them, please refer to [the llama.cpp documentation](https://github.com/ggerganov/llama.cpp/blob/master/examples/main/README.md)
## How to run in `text-generation-webui`
Further instructions can be found in the text-generation-webui documentation, here: [text-generation-webui/docs/04 ‐ Model Tab.md](https://github.com/oobabooga/text-generation-webui/blob/main/docs/04%20%E2%80%90%20Model%20Tab.md#llamacpp).
## How to run from Python code
You can use GGUF models from Python using the [llama-cpp-python](https://github.com/abetlen/llama-cpp-python) or [ctransformers](https://github.com/marella/ctransformers) libraries. Note that at the time of writing (Nov 27th 2023), ctransformers has not been updated for some time and is not compatible with some recent models. Therefore I recommend you use llama-cpp-python.
### How to load this model in Python code, using llama-cpp-python
For full documentation, please see: [llama-cpp-python docs](https://abetlen.github.io/llama-cpp-python/).
#### First install the package
Run one of the following commands, according to your system:
```shell
# Base ctransformers with no GPU acceleration
pip install llama-cpp-python
# With NVidia CUDA acceleration
CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python
# Or with OpenBLAS acceleration
CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python
# Or with CLBLast acceleration
CMAKE_ARGS="-DLLAMA_CLBLAST=on" pip install llama-cpp-python
# Or with AMD ROCm GPU acceleration (Linux only)
CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python
# Or with Metal GPU acceleration for macOS systems only
CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python
# In windows, to set the variables CMAKE_ARGS in PowerShell, follow this format; eg for NVidia CUDA:
$env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on"
pip install llama-cpp-python
```
#### Simple llama-cpp-python example code
```python
from llama_cpp import Llama
# Set gpu_layers to the number of layers to offload to GPU. Set to 0 if no GPU acceleration is available on your system.
llm = Llama(
model_path="./meditron-70b.Q4_K_M.gguf", # Download the model file first
n_ctx=4096, # The max sequence length to use - note that longer sequence lengths require much more resources
n_threads=8, # The number of CPU threads to use, tailor to your system and the resulting performance
n_gpu_layers=35 # The number of layers to offload to GPU, if you have GPU acceleration available
)
# Simple inference example
output = llm(
"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant", # Prompt
max_tokens=512, # Generate up to 512 tokens
stop=[""], # Example stop token - not necessarily correct for this specific model! Please check before using.
echo=True # Whether to echo the prompt
)
# Chat Completion API
llm = Llama(model_path="./meditron-70b.Q4_K_M.gguf", chat_format="llama-2") # Set chat_format according to the model you are using
llm.create_chat_completion(
messages = [
{"role": "system", "content": "You are a story writing assistant."},
{
"role": "user",
"content": "Write a story about llamas."
}
]
)
```
## How to use with LangChain
Here are guides on using llama-cpp-python and ctransformers with LangChain:
* [LangChain + llama-cpp-python](https://python.langchain.com/docs/integrations/llms/llamacpp)
* [LangChain + ctransformers](https://python.langchain.com/docs/integrations/providers/ctransformers)
## Discord
For further support, and discussions on these models and AI in general, join us at:
[TheBloke AI's Discord server](https://discord.gg/theblokeai)
## Thanks, and how to contribute
Thanks to the [chirper.ai](https://chirper.ai) team!
Thanks to Clay from [gpus.llm-utils.org](llm-utils)!
I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.
If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.
Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
* Patreon: https://patreon.com/TheBlokeAI
* Ko-Fi: https://ko-fi.com/TheBlokeAI
**Special thanks to**: Aemon Algiz.
**Patreon special mentions**: Brandon Frisco, LangChain4j, Spiking Neurons AB, transmissions 11, Joseph William Delisle, Nitin Borwankar, Willem Michiel, Michael Dempsey, vamX, Jeffrey Morgan, zynix, jjj, Omer Bin Jawed, Sean Connelly, jinyuan sun, Jeromy Smith, Shadi, Pawan Osman, Chadd, Elijah Stavena, Illia Dulskyi, Sebastain Graf, Stephen Murray, terasurfer, Edmond Seymore, Celu Ramasamy, Mandus, Alex, biorpg, Ajan Kanaga, Clay Pascal, Raven Klaugh, 阿明, K, ya boyyy, usrbinkat, Alicia Loh, John Villwock, ReadyPlayerEmma, Chris Smitley, Cap'n Zoog, fincy, GodLy, S_X, sidney chen, Cory Kujawski, OG, Mano Prime, AzureBlack, Pieter, Kalila, Spencer Kim, Tom X Nguyen, Stanislav Ovsiannikov, Michael Levine, Andrey, Trailburnt, Vadim, Enrico Ros, Talal Aujan, Brandon Phillips, Jack West, Eugene Pentland, Michael Davis, Will Dee, webtim, Jonathan Leane, Alps Aficionado, Rooh Singh, Tiffany J. Kim, theTransient, Luke @flexchar, Elle, Caitlyn Gatomon, Ari Malik, subjectnull, Johann-Peter Hartmann, Trenton Dambrowitz, Imad Khwaja, Asp the Wyvern, Emad Mostaque, Rainer Wilmers, Alexandros Triantafyllidis, Nicholas, Pedro Madruga, SuperWojo, Harry Royden McLaughlin, James Bentley, Olakabola, David Ziegler, Ai Maven, Jeff Scroggin, Nikolai Manek, Deo Leter, Matthew Berman, Fen Risland, Ken Nordquist, Manuel Alberto Morcote, Luke Pendergrass, TL, Fred von Graf, Randy H, Dan Guido, NimbleBox.ai, Vitor Caleffi, Gabriel Tamborski, knownsqashed, Lone Striker, Erik Bjäreholt, John Detwiler, Leonard Tan, Iucharbius
Thank you to all my generous patrons and donaters!
And thank you again to a16z for their generous grant.
# Original model card: EPFL LLM Team's Meditron 70B
# Model Card for Meditron-70B-v1.0
Meditron is a suite of open-source medical Large Language Models (LLMs).
Meditron-70B is a 70 billion parameters model adapted to the medical domain from Llama-2-70B through continued pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, a [new dataset](https://huggingface.co/datasets/epfl-llm/guidelines) of internationally-recognized medical guidelines, and general domain data from [RedPajama-v1](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T).
Meditron-70B, finetuned on relevant training data, outperforms Llama-2-70B, GPT-3.5 (`text-davinci-003`, 8-shot), and Flan-PaLM on multiple medical reasoning tasks.
Advisory Notice
While Meditron is designed to encode medical knowledge from sources of high-quality evidence, it is not yet adapted to deliver this knowledge appropriately, safely, or within professional actionable constraints.
We recommend against deploying Meditron in medical applications without extensive use-case alignment, as well as additional testing, specifically including randomized controlled trials in real-world practice settings.
## Model Details
- **Developed by:** [EPFL LLM Team](https://huggingface.co/epfl-llm)
- **Model type:** Causal decoder-only transformer language model
- **Language(s):** English (mainly)
- **Model License:** [LLAMA 2 COMMUNITY LICENSE AGREEMENT](https://huggingface.co/meta-llama/Llama-2-70b/raw/main/LICENSE.txt)
- **Code License:** [APACHE 2.0 LICENSE](LICENSE)
- **Continue-pretrained from model:** [Llama-2-70B](https://huggingface.co/meta-llama/Llama-2-70b)
- **Context length:** 4K tokens
- **Input:** Text-only data
- **Output:** Model generates text only
- **Status:** This is a static model trained on an offline dataset. Future versions of the tuned models will be released as we enhance model's performance.
- **Knowledge Cutoff:** August 2023
### Model Sources
- **Repository:** [epflLLM/meditron](https://github.com/epfLLM/meditron)
- **Trainer:** [epflLLM/Megatron-LLM](https://github.com/epfLLM/Megatron-LLM)
- **Paper:** *[MediTron-70B: Scaling Medical Pretraining for Large Language Models](https://arxiv.org/abs/2311.16079)*
## Uses
Meditron-70B is being made available for further testing and assessment as an AI assistant to enhance clinical decision-making and enhance access to an LLM for healthcare use. Potential use cases may include but are not limited to:
- Medical exam question answering
- Supporting differential diagnosis
- Disease information (symptoms, cause, treatment) query
- General health information query
### Direct Use
It is possible to use this model to generate text, which is useful for experimentation and understanding its capabilities.
It should not be used directly for production or work that may impact people.
### Downstream Use
Meditron-70B is a foundation model that can be finetuned, instruction-tuned, or RLHF-tuned for specific downstream tasks and applications.
The main way we have used this model is finetuning for downstream question-answering tasks, but we encourage using this model for additional applications.
Specific formatting needs to be followed to prompt our finetuned models, including the `<|im_start|>`, `<|im_end|>` tags, and `system`, `question`, `answer` identifiers.
"""
<|im_start|>system
{system_message}<|im_end|>
<|im_start|>question
{prompt}<|im_end|>
<|im_start|>answer
"""
**Note 1**: The above formatting is not required for running the base model (this repository)
**Note 2**: the above formatting is just an example of a finetuning template. This format is not a requirement if you use your own formatting option for the finetuning of the model.
To run proper generation with this base model, we recommend using a high-throughput and memory-efficient inference engine, such as [vLLM](https://github.com/vllm-project/vllm), with a UI that supports chat and text generation, such as [BetterChatGPT](https://github.com/ztjhz/BetterChatGPT)
To see more details about model deployment and generation, please see our [documentation](https://github.com/epfLLM/meditron/blob/main/deployment/README.md).
### Out-of-Scope Use
We do not recommend using this model for natural language generation in a production environment, finetuned or otherwise.
## Truthfulness, Helpfulness, Risk, and Bias
We did an initial assessment of Meditron models' **Truthfulness** against baseline models and consumer-level medical models.
We use TruthfulQA (multiple choice) as the main evaluation benchmark.
We only focus on the categories that are relevant to the medical domain, including Health, Nutrition, Psychology, and Science.
For 7B models, we perform one-shot evaluations for consistent answer generation.
For 70B models, the evaluations are under the zero-shot setting.
Below, we report the detailed truthfulness performance of each category.
| | | | | | | | |
| --- | ------ |----- |----- |----- |----- |----- |----- |
|Category | meditron-70b | llama-2-70b | med42-70b* | meditron-7b | llama-2-7b | PMC-llama-7b |
|Health | 81.8 | 69.1 | 83.6 | 27.3 | 16.4 | 3.6 |
|Nutrition | 77.9 | 68.8 | 62.5 | 31.1 | 12.5 | 6.3 |
|Psychology| 47.4 | 36.8 | 52.6 | 21.1 | 10.5 | 0.0 |
|Science | 77.8 | 44.4 | 33.3 | 33.3 | 11.1 | 0.0 |
|Avg | 71.2 | 54.8 | 58.0 | 28.3 | 12.6 | 2.5 |
| | | | | | | |
For a more detailed performance analysis, please see our paper.
For **Helpfulness**, **Risk** and **Bias**, we provide a comprehensive qualitative generation report of Meditron-70B on queries designed by medical experts.
Each query targets specific aspects of helpfulness (medical accuracy, up-to-date information, etc.), risk (public health, medical ethics, etc.) and bias (gender, age, race, etc.).
Please see the detailed generations in our paper. We compare our generations to Llama-2-70B and ChatGPT-3.5 (version Nov, 27, 2023)
Significant research is still required to fully explore potential bias, fairness, and safety issues with this language model.
### Recommendations
**IMPORTANT!**
Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model.
While this model is capable of generating natural language text, we have only begun to explore this capability and its limitations.
Understanding these limitations is especially important in a domain like medicine.
Therefore, we strongly recommend against using this model in production for natural language generation or for professional purposes related to health and medicine without comprehensive testing for your application.
## Training Details
### Training Data
Meditron’s domain-adaptive pre-training corpus GAP-Replay combines 48.1B tokens from four corpora:
- [**Clinical Guidelines**](https://huggingface.co/datasets/epfl-llm/guidelines): a new dataset of 46K internationally-recognized clinical practice guidelines from various healthcare-related sources, including hospitals and international organizations.
- **Medical Paper Abstracts**: 16.1M abstracts extracted from closed-access PubMed and PubMed Central papers.
- **Medical Papers**: full-text articles extracted from 5M publicly available PubMed and PubMed Central papers.
- **Replay Data**: 400M tokens of general domain pretraining data sampled from [RedPajama-v1](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T)
#### Data Preprocessing
Please see the detailed preprocessing procedure in our paper.
### Training Procedure
We used the [Megatron-LLM](https://github.com/epfLLM/Megatron-LLM) distributed training library, a derivative of Nvidia's Megatron LM project, to optimize training efficiency.
Hardware consists of 16 nodes of 8x NVIDIA A100 (80GB) SXM GPUs connected by NVLink and NVSwitch with a single Nvidia ConnectX-6 DX network card and equipped with 2 x AMD EPYC 7543 32-Core Processors and 512 GB of RAM.
The nodes are connected via RDMA over Converged Ethernet.
Our three-way parallelism scheme uses:
- Data Parallelism (DP -- different GPUs process different subsets of the batches) of 2,
- Pipeline Parallelism (PP -- different GPUs process different layers) of 8,
- Tensor Parallelism (TP -- different GPUs process different subtensors for matrix multiplication) of 8.
#### Training Hyperparameters
| | |
| --- | ------ |
| bf16 | true |
| lr | 1.5e-4 |
| eps | 1e-5 |
| betas | \[0.9, 0.95\] |
| clip_grad | 1 |
| weight decay | 0.1 |
| DP size | 2 |
| TP size | 8 |
| PP size | 8 |
| seq length | 4096 |
| lr scheduler | cosine|
| min lr | 1e-6 |
| warmup iteration | 2000 |
| micro batch size | 2 |
| global batch size | 512 |
| | |
#### Speeds, Sizes, Times
The model was trained in September and October 2023.
The model architecture is exactly Llama 2, meaning
| | |
| --- | ------ |
| Model size | 70B |
| Hidden dimension | 8192 |
| Num. attention heads | 64 |
| Num. layers | 80 |
| | | |
We train the 70B model on 48e9 tokens, at a throughput of about 40,200 tokens / second.
This amounts to a bfloat16 model flops utilization of roughly 42.3\%.
## Evaluation
### Testing Data & Metrics
#### Testing Data
- [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)
#### Metrics
- Accuracy: suite the evaluation of multiple-choice question-answering tasks.
### Results
We finetune meditron-70b and llama-2-70b on each benchmark (pubmedqa, medmcqa, medqa)'s training data individually.
We report the finetuned models' performance with self-consistency chain-of-thought as the inference mode.
For MMLU-Medical, models finetuned on MedMCQA are used for inference.
For MedQA-4-Option, models finetuned on MedQA are used for inference.
For a more detailed performance analysis, please see our paper.
| | | | | | |
| --- | ------ |----- |----- |----- |----- |
|Dataset| meditron-70b | llama-2-70b | med42-70b* | clinical-camel-70b* |
|MMLU-Medical | 77.6 | 77.9 | 74.5 | 65.7 |
|PubMedQA | 81.6 | 80.0 | 61.2 | 67.0 |
|MedMCQA | 66.0 | 62.6 | 59.2 | 46.7 |
|MedQA | 64.4 | 61.5 | 59.1 | 50.8 |
|MedQA-4-Option| 70.2 | 63.8 | 63.9 | 56.8 |
|Avg | 72.0 | 69.2 | 63.6 | 57.4 |
| | | | | | |
**Note**: models with * are already instruction-tuned, so we exclude them from further finetuning on any training data.
## Environmental Impact
- **Hardware Type:** 128 x NVIDIA A100 (80GB) SXM
- **Total GPU hours:** 42,496
- **Hardware Provider:** EPFL Research Computing Platform
- **Compute Region:** Switzerland
- **Carbon Emitted:** Switzerland has a carbon efficiency of 0.016 kgCO2/kWh (https://www.carbonfootprint.com/docs/2018_8_electricity_factors_august_2018_-_online_sources.pdf). 332 hours of 128 A100s means 42496 hours at a TDP of 400W. Assuming a Power Usage effectiveness of 1.8, total emissions are estimated to be:
(400W / 1000W/kWh / GPU * 0.016 kgCO2/kWh * 332 h * 128 GPU) * 1.8 PUE = 486 kgCO2.
## Citation
**BibTeX:**
If you use Meditron or its training data, please cite our work:
```
@misc{chen2023meditron70b,
title={MEDITRON-70B: Scaling Medical Pretraining for Large Language Models},
author={Zeming Chen and Alejandro Hernández-Cano and Angelika Romanou and Antoine Bonnet and Kyle Matoba and Francesco Salvi and Matteo Pagliardini and Simin Fan and Andreas Köpf and Amirkeivan Mohtashami and Alexandre Sallinen and Alireza Sakhaeirad and Vinitra Swamy and Igor Krawczuk and Deniz Bayazit and Axel Marmet and Syrielle Montariol and Mary-Anne Hartley and Martin Jaggi and Antoine Bosselut},
year={2023},
eprint={2311.16079},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@software{epfmedtrn,
author = {Zeming Chen and Alejandro Hernández Cano and Angelika Romanou and Antoine Bonnet and Kyle Matoba and Francesco Salvi and Matteo Pagliardini and Simin Fan and Andreas Köpf and Amirkeivan Mohtashami and Alexandre Sallinen and Alireza Sakhaeirad and Vinitra Swamy and Igor Krawczuk and Deniz Bayazit and Axel Marmet and Syrielle Montariol and Mary-Anne Hartley and Martin Jaggi and Antoine Bosselut},
title = {MediTron-70B: Scaling Medical Pretraining for Large Language Models},
month = November,
year = 2023,
url = {https://github.com/epfLLM/meditron}
}
```