mistral-finetune is a light-weight codebase that enables memory-efficient and performant finetuning of Mistral's models. It is based on LoRA, a training paradigm where most weights are frozen and only 1-2% of additional weights in the form of low-rank matrix perturbations are trained.

For maximum efficiency it is recommended to use an A100 or H100 GPU. The codebase is optimized for multi-GPU-single-node training setups, but for smaller models, such as the 7B a single GPU suffices.

Note

• The goal of this repository is to provide a simple, guided entrypoint to finetune Mistral models. As such, it is fairly opinionated (especially around data formatting) and does not aim at being exhaustive across multiple model architectures or hardware types. For more generic approaches, you can check out some other great projects like torchtune.

• 19.07.2024: mistral-finetune is now compatible with Mistral Nemo!
  • 1. Download the new checkpoints here and set model_id_or_path to the new checkpoint
  • 2. Run pip install --upgrade mistral-common to have a version that supports the Tekkenizer (>=1.3.1).
  • 3. Fine-tuning Mistral-Nemo requires currently much more memory due to a larger vocabulary size which spikes the peak memory requirement of the CE loss (we'll soon add an improved CE loss here). For now set seq_len to 16384 or 8192
  • 4. It is recommended to use the same hyperparameters as for the 7B v3.

To get started with Mistral LoRA fine-tuning, follow these steps:

1. Clone this repository:

cd $HOME && git clone https://github.com/mistralai/mistral-finetune.git

2. Install all required dependencies:

cd mistral-finetune
pip install -r requirements.txt

We recommend fine-tuning one of the official Mistral models which you can download here:

Important Notice: For 8x7B Base V1 and 8x7B Instruct V1, it is necessary to use our v3 tokenizer and extend the vocabulary size to 32768 prior to fine-tuning. For detailed instructions on this process, please refer to the "Model extension" section.

E.g., to download the 7B-base model you can run the following command:

mkdir -p ~/${HOME}/mistral_models
cd ${HOME} && wget https://models.mistralcdn.com/mistral-7b-v0-3/mistral-7B-v0.3.tar
tar -xf mistral-7B-v0.3.tar -C mistral_models

Make sure to modify your training script and add the path to the downloaded folder as model_id_or_path.

E.g., modify example/7B.yaml to include the absolute path to $HOME/mistral_models/7B:

model_id_or_path: "/Users/johndoe/mistral_models/7B"

To ensure effective training, mistral-finetune has strict requirements for how the training data has to be formatted.

All data files must be stored in jsonl format files.

You can build two types of data files:

Pretrain data corresponds to plain text data stored in the "text" key. E.g:

{"text": "Text contained in document n°1"}
{"text": "Text contained in document n°2"}

Currently two different types of instruction following data are supported:

• Instruct: conversational data stored in the "messages" key in the form of a list. Each list item is a dictionary containing the "content" and "role" keys. "role" is a string being one of "user", "assistant" or "system". The loss will only be computed if "role" == "assistant". E.g.:

{
  "messages": [
    {
      "role": "user",
      "content": "User interaction n°1 contained in document n°1"
    },
    {
      "role": "assistant",
      "content": "Bot interaction n°1 contained in document n°1"
    },
    {
      "role": "user",
      "content": "User interaction n°2 contained in document n°1"
    },
    {
      "role": "assistant",
      "content": "Bot interaction n°2 contained in document n°1"
    }
  ]
}
{
  "messages": [
    {
      "role": "user",
      "content": "User interaction n°1 contained in document n°2"
    },
    {
      "role": "assistant",
      "content": "Bot interaction n°1 contained in document n°2"
    },
    {
      "role": "user",
      "content": "User interaction n°2 contained in document n°2"
    },
    {
      "role": "assistant",
      "content": "Bot interaction n°2 contained in document n°2",
      "weight": 0,  # don't train on n°2
    },
    {
      "role": "user",
      "content": "User interaction n°3 contained in document n°2"
    },
    {
      "role": "assistant",
      "content": "Bot interaction n°3 contained in document n°2"
    }
  ]
}

• Function calling: conversational data stored in the "messages" key in the form of a list. Each list item is a dictionary containing the "role" and "content" or "tool_calls" keys. "role" is a string being one of "user", "assistant", "system", or "tool". The loss will only be computed if "role" == "assistant".

Note: In function calling the "id" of "tool_calls" and the "tool_call_id" are randomly generated strings of exactly 9 chars. We recommend to generate this automatically in a data preparation script as is done here.

E.g.:

{
  "messages": [
    {
      "role": "system",
      "content": "You are a helpful assistant who has access to the following functions to help the user, you can use the functions if needed"
    },
    {
      "role": "user",
      "content": "Can you help me generate an anagram of the word \"listen\"?"
    },
    {
      "role": "assistant",
      "tool_calls": [
        {
          "id": "TX92Jm8Zi",
          "type": "function",
          "function": {
            "name": "generate_anagram",
            "arguments": "{\"word\": \"listen\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "{\"anagram\": \"silent\"}",
      "tool_call_id": "TX92Jm8Zi"
    },
    {
      "role": "assistant",
      "content": "The anagram of the word \"listen\" is \"silent\"."
    },
    {
      "role": "user",
      "content": "That's amazing! Can you generate an anagram for the word \"race\"?"
    },
    {
      "role": "assistant",
      "tool_calls": [
        {
          "id": "3XhQnxLsT",
          "type": "function",
          "function": {
            "name": "generate_anagram",
            "arguments": "{\"word\": \"race\"}"
          }
        }
      ]
    }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "generate_anagram",
        "description": "Generate an anagram of a given word",
        "parameters": {
          "type": "object",
          "properties": {
            "word": {
              "type": "string",
              "description": "The word to generate an anagram of"
            }
          },
          "required": [
            "word"
          ]
        }
      }
    }
  ]
}

Before starting a training run you should verify that your dataset is correctly formatted and get an estimation of the training time. You can do so by using the ./utils/validate_data script.

Note that this step is crucial to ensure that the data is correctly formatted.

Let's go over a simple example to train a model in instruction following:

• 1. Load a chunk of Ultachat_200k

Create the data folder and navigate to the folder.

cd $HOME && mkdir -p data && cd $HOME/data

Load the data into a Pandas Dataframe.

Note: Make sure to have pandas and pyarrow installed (pip install pandas pyarrow).

import pandas as pd

df = pd.read_parquet('https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k/resolve/main/data/test_gen-00000-of-00001-3d4cd8309148a71f.parquet')

• 2. Split into train and eval

df_train=df.sample(frac=0.95,random_state=200)
df_eval=df.drop(df_train.index)

• 3. Save data to jsonl

df_train.to_json("ultrachat_chunk_train.jsonl", orient="records", lines=True)
df_eval.to_json("ultrachat_chunk_eval.jsonl", orient="records", lines=True)

• 4. Modify your training yaml to include the ultrachat dataset and verify the yaml

Modify example/7B.yaml to include the absolute path to $HOME/data/ultrachat_chunk_train.jsonl as well as a dataset mixing weight for training and $HOME/data/ultrachat_chunk_eval.jsonl for eval, e.g.

data:
  instruct_data: "/Users/johndoe/data/ultrachat_chunk_train.jsonl"
  eval_instruct_data: "/Users/johndoe/data/ultrachat_chunk_eval.jsonl"

Now you can verify your training yaml to make sure the data is correctly formatted and to get an estimate of your training time.

cd $HOME/mistral-finetune
python -m utils.validate_data --train_yaml example/7B.yaml

Upon completion you should see an error report with many of the following errors:

The data in line 1412 of dataset /Users/johndoe/data/ultrachat_chunk_eval.jsonl is incorrectly formatted. Expected last role to be one of: [assistant] but got user
The data in line 1413 of dataset /Users/johndoe/data/ultrachat_chunk_eval.jsonl is incorrectly formatted. Expected last role to be one of: [assistant] but got user
The data in line 1414 of dataset /Users/johndoe/data/ultrachat_chunk_eval.jsonl is incorrectly formatted. Expected last role to be one of: [assistant] but got user
The data in line 1415 of dataset /Users/johndoe/data/ultrachat_chunk_eval.jsonl is incorrectly formatted. Expected last role to be one of: [assistant] but got user

Many conversations seem to end with the 'user' role which is unnecessary as we only train on 'assistant' messages and thus would unnecessarily process data.

You can make use of ./utils/reformat_data.py to correct the data:

cd $HOME/mistral-finetune
python -m utils.reformat_data $HOME/data/ultrachat_chunk_train.jsonl
python -m utils.reformat_data $HOME/data/ultrachat_chunk_eval.jsonl

You should see that a couple of samples will be skipped.

• 5. Potentially change number of training steps

Upon correction of the dataset, run the script again

cd $HOME/mistral-finetune
python -m utils.validate_data --train_yaml example/7B.yaml

You should get a summary of the data input and training parameters:

Train States
 --------------------
{
   "expected": {
       "eta": "00:52:44",
       "data_tokens": 25169147,
       "train_tokens": 131072000,
       "epochs": "5.21",
       "max_steps": 500,
       "data_tokens_per_dataset": {
           "/Users/johndoe/data/ultrachat_chunk_train.jsonl": "25169147.0"
       },
       "train_tokens_per_dataset": {
           "/Users/johndoe/data/ultrachat_chunk_train.jsonl": "131072000.0"
       },
       "epochs_per_dataset": {
           "/Users/johndoe/data/ultrachat_chunk_train.jsonl": "5.2"
       }
   },
}

Having max_steps set to 500 would lead to iterating through the dataset roughly 5 times which is reasonable, but might be a bit too much. A recommended setting is shown below which would only take 30min on a 8xH100 cluster.

Next let's go over a more advanced use case to fine-tune a model on function calling. Function calling requires the data to be in the format as explained above. Let's go over an example.

• 1. Load a chat-formatted version of the Glaive function calling dataset

Create the data folder and navigate to the folder.

cd $HOME && mkdir -p data && cd $HOME/data

Load the data into a Pandas Dataframe.

Note: Make sure to have pandas and pyarrow installed (pip install pandas pyarrow).

import pandas as pd

df = pd.read_parquet('https://huggingface.co/datasets/Locutusque/function-calling-chatml/resolve/main/data/train-00000-of-00001-f0b56c6983b4a78f.parquet')

• 2. Split into train and eval

df_train=df.sample(frac=0.95,random_state=200)
df_eval=df.drop(df_train.index)

• 3. Save data to jsonl

df_train.to_json("glaive_train.jsonl", orient="records", lines=True)
df_eval.to_json("glaive_eval.jsonl", orient="records", lines=True)

• 4. Reformat dataset

As one can see the dataset does not follow the required function calling format, so it will need to be reformatted. Among other things "from" should be renamed to "user" and superfluous "\n" characters should be removed. For this dataset you can make use of ./utils/reformat_data_glaive.py:

cd $HOME/mistral-finetune
python -m utils.reformat_data_glaive $HOME/data/glaive_train.jsonl
python -m utils.reformat_data_glaive $HOME/data/glaive_eval.jsonl

Running this command will make sure that most samples are in the correct format.

Note: It is impossible to write reformatting scripts that work for all kinds of datasets. If you have datasets that don't yet follow the required format above, you will most probably have to create a reformatting script yourself (mistral-chat or chat-gpt is your best friend here!).

• 5. Validate dataset

You can now validate the dataset by setting data.instruct_data and data.eval_instruct_data to $HOME/data/glaive_train.jsonl and $HOME/data/glaive_eval.jsonl in example/7B.yaml respectively.

The reformatted datasets still have some errors which can be removed with --create_corrected. For this, make sure to add --create_corrected as follows:

cd $HOME/mistral-finetune
python -m utils.validate_data --train_yaml example/7B.yaml --create_corrected

Running this command will show a couple of errors and save two new datasets $HOME/data/glaive_train.jsonl.corrected and $HOME/data/glaive_eval.jsonl.corrected. Make sure to use these two dataset in example/7B.yaml and run the command again. Now the dataset should be correctly formatted!

Having followed the dataset verification section, we can now start training. For faster training, we recommend setting max_steps to only 300. Make sure to define run_dir to your experiment folder and optionally set wandb_project to a Weights & Biases project for logging`, e.g.:

max_steps: 300
run_dir: "/Users/johndoe/ultra_chat_test"
wandb.project: ultra_chat

Optionally you can also set wandb

Save the training configuration and start training! Make sure to set --nproc-per-node to the number of available GPUs.

cd $HOME/mistral-finetune
torchrun --nproc-per-node 8 --master_port $RANDOM -m train example/7B.yaml

Training on ultra-chat should take around 30min on a 8xH100 node and the resulting weights should give an MT Bench score around 6.3.

Training on glaive should take around 1h on a 8xH100 node and the resulting weights should work nicely for function calling.

The example mistral-finetune/examples/7B defines reasonable parameters for learning rate, weight decay, etc... but you are advised to customize these settings for your use case.

Generally, a training configuration should fill the following parameters:

• model_id_or_path defines the model to start training from. This can be a path to a pre-trained model or a local model directory.
• run_dir defines the directory where training checkpoints and metrics are stored.
• seq_len defines the sequence length for training. This is the maximum length of input sequences the model will process. Samples are packed to reach a length of seq_len for maximum training efficiency.
• batch_size defines the number of training examples used per GPU. Note: The overall effective batch_size (in tokens) across all GPUs equals num_gpus x batch_size x seq_len.
• max_steps defines the maximum number of training steps. This is the total number of iterations the training process will run. It can be adjusted based on the specific needs of your training scenario. Total number of tokens seen during training is max_steps x num_gpus x batch_size x seq_len.
• optim.lr defines the learning rate. This is the initial learning rate for the optimizer.
• optim.weight_decay defines weight decay. Weight decay is a regularization technique used to prevent overfitting by penalizing large weights. We recommend leaving it at 0.1.
• optim.pct_start defines the percentage of the total training steps used for the learning rate warm-up phase before it starts to decrease. It corresponds to pct_start of PyTorch's OneCycleLR.
• lora.rank defines the size of the LoRA (Low-Rank Adaptation) adapters. We recommend 64 or less, which adjusts the rank of the low-rank decomposition used in LoRA.
• seed defines the random seed for initialization and data shuffling/sampling. Setting a seed ensures reproducibility of results.
• log_freq defines the logging frequency. This specifies how often (in steps) to log training metrics.
• data.instruct_data is the path to the instruction data used for training. This field has to be filled with one or multiple data sources in the format as explained above. Each data source should either be a path to a jsonl file or a path to a directory containing jsonl files followed by a weighting to define the importance of this dataset: <path/to/data_source>:<weight>. E.g.: data.instruct_data: "/path/to/data1.jsonl:5.,/path/to/data2.jsonl:1.,/path/to/dir_of_jsonls:1."
• data.data is an optional path to additional pretraining data in the format as explained above. Note that this field can be left blank.
• data.eval_instruct_data is an optional path to evaluation instruction data to run cross-validation at every eval_freq steps. Cross-validation metrics are displayed as loss and perplexity.
• eval_freq defines how often (in steps) to evaluate the model. This specifies the interval at which the model is evaluated on the validation set.
• no_eval is a flag to enable or disable intermediate evaluation. Setting it to False enables periodic evaluation during training.
• ckpt_freq defines how often (in steps) to save checkpoints. This specifies the interval at which the model's state is saved.
• save_adapters defines whether to only save the trained LoRA checkpoints or whether the trained LoRA should directly be merged into the base model and saved. Note: When setting save_adapters=False make sure that you have enough CPU and GPU memory to save the full model on a single process (this is usually only possible for the 7B model).
• wandb.key is used to pass your Weights & Biases (wandb) API key for logging. This allows you to log training metrics to the wandb dashboard.
• wandb.project defines the wandb project name. This is where the training run will be logged in the wandb interface.

Once your model is trained, you should try it out in inference. We recommend using mistral-inference.

Make sure to have mistral_inference correctly installed:

pip install mistral_inference

Assuming your lora.safetensors is saved under $HOME/ultra_chat_test/checkpoints/checkpoint_000300/consolidated/lora.safetensors, you can chat with the model using mistral_inference, e.g.:

mistral-chat /mnt/slow/runs/patrick/mistral-finetune/7B/ --max_tokens 256 --temperature 1.0 --instruct --lora_path $HOME/ultra_chat_test/checkpoints/checkpoint_000300/consolidated/lora.safetensors

We have added explicit support for Weights and Biases to help you monitor and visualize your training runs. This integration allows you to log various metrics and track experiments easily.

To use Weights and Biases with mistral-finetune, follow these steps:

1. Install Weights and Biases:

   Make sure you have the wandb library installed. You can install it using pip:

   pip install wandb

Once the training starts, you can monitor the progress in real-time by visiting your wandb project dashboard. All metrics, including training loss, evaluation loss, learning rate, etc., will be logged and visualized.

For more details on how to use wandb, visit the Weights and Biases documentation.

Important: Note that one can only fine-tune mistral models that are compatible with the v3 tokenizer which entails that the models have a vocabulary size of 32768 - not 32000. One can however easily extend older version of vocabulary size 32000 to have a vocabulary size of 32768 by using:

python -m utils.extend_model_vocab --original_model_ckpt /folder/to/old/model --extended_model_ckpt /folder/to/extended/model

Once the extension has worked, one can fine-tune using the newly created model checkpoint in /folder/to/extended/model.

• What's the best practice of fine-tuning MoEs?

We see a higher degree of performance variance in when fine-tuning MoE models. It's not unusual to find that fine-tuning MoE models with different seeds can lead to a high variance in performance. We did not observe such a high variance with dense models. Therefore, we suggest running multiple instances of the same fine-tuning process on MoEs models and selecting the one that performs best.

• How can I determine the number of tokens used during the model training process?

You can use the following script to find out: https://github.com/mistralai/mistral-finetune/blob/main/utils/validate_data.py. This script accepts a .yaml training file as input and returns the number of tokens the model is being trained on.

• What should I do if I encounter a CUDA out-of-memory error?

One possible solution is to reduce the batch size per GPU. The batch size is equal to seq_len x batch_size. Try setting batch_size to 1 and reduce seq_len. You can define the batch_size and seq_len in the .yaml file.