---
license: apache-2.0
datasets:
- CohereForAI/aya_dataset
- argilla/databricks-dolly-15k-curated-multilingual
- Gael540/dataSet_ens_sup_fr-v1
- ai2-adapt-dev/flan_v2_converted
- OpenAssistant/oasst1
language:
- fr
- en
- de
- it
- es
base_model:
- OpenLLM-France/Lucie-7B
pipeline_tag: text-generation
---

# Model Card for Lucie-7B-Instruct-human-data

* [Model Description](#model-description)
<!-- * [Uses](#uses) -->
* [Training Details](#training-details)
  * [Training Data](#training-data)
  * [Preprocessing](#preprocessing)
  * [Instruction template](#instruction-template)
  * [Training Procedure](#training-procedure)
<!-- * [Evaluation](#evaluation) -->
* [Testing the model](#testing-the-model)
  * [Test with ollama](#test-with-ollama)
  * [Test with vLLM](#test-with-vllm)
* [Citation](#citation)
* [Acknowledgements](#acknowledgements)
* [Contact](#contact)

## Model Description

Lucie-7B-Instruct-human-data is a fine-tuned version of [Lucie-7B](), an open-source, multilingual causal language model created by OpenLLM-France.

Lucie-7B-Instruct-human-data is fine-tuned on human-produced instructions collected either from open annotation campaigns or by applying templates to extant datasets. The performance of Lucie-7B-Instruct-human-data falls below that of [Lucie-7B-Instruct](https://huggingface.co/OpenLLM-France/Lucie-7B-Instruct); the interest of the model is to show what can be done to fine-tune LLMs to follow instructions without appealing to third party LLMs.

Note that both Lucie-7B-Instruct-human-data and Lucie-7B-Instruct are optimized for generation of French text. They have not been trained for code generation or optimized for math. Such capacities can be improved through further fine-tuning and alignment with methods such as DPO, RLHF, etc.

While Lucie-7B-Instruct-human-data is trained on sequences of 4096 tokens, its base model, Lucie-7B has a context size of 32K tokens. Based on Needle-in-a-haystack evaluations, Lucie-7B-Instruct-human-data maintains the capacity of the base model to handle 32K-size context windows.

## Training details
### Training data

Lucie-7B-Instruct-human-data is trained on the following datasets published by third parties:
* [Aya Dataset](https://huggingface.co/datasets/CohereForAI/aya_dataset) (English, 3944 samples; French, 1422; German, 241; Italian, 738; Spanish, 3854)
* [Dolly](https://huggingface.co/datasets/argilla/databricks-dolly-15k-curated-multilingual) (English, French, German, Spanish; 15015 x 4 samples)
* [ENS](https://huggingface.co/datasets/Gael540/dataSet_ens_sup_fr-v1) (French, 394 samples)
* [FLAN v2 Converted](https://huggingface.co/datasets/ai2-adapt-dev/flan_v2_converted) (English, 78580 samples)
* [Open Assistant 1](https://huggingface.co/datasets/OpenAssistant/oasst1) (English, 21151 samples; French, 1223; German, 1515; Italian, 370; Spanish, 14078)
* [Oracle](https://github.com/opinionscience/InstructionFr/tree/main/wikipedia) (French, 4613 samples)
* [PIAF](https://www.data.gouv.fr/fr/datasets/piaf-le-dataset-francophone-de-questions-reponses/) (French, 1849 samples)


And the following datasets developed for the Lucie instruct models:
* [Croissant Aligned Instruct](https://huggingface.co/datasets/OpenLLM-France/Croissant-Aligned-Instruct) (French-English, 20K examples sampled randomly from 80K total)
* Hard-coded prompts concerning OpenLLM and Lucie (based on [allenai/tulu-3-hard-coded-10x](https://huggingface.co/datasets/allenai/tulu-3-hard-coded-10x))
    * French: openllm_french.jsonl (24x10 samples)
    * English: openllm_english.jsonl (24x10 samples)

### Preprocessing
* Filtering by language: Aya Dataset, Dolly and Open Assistant were filtered to keep only languages on which Lucie-7B was trained.
* Filtering by keyword: Examples containing assistant responses were filtered out from Open Assistant if the responses contained a keyword from the list [filter_strings](https://github.com/OpenLLM-France/Lucie-Training/blob/98792a1a9015dcf613ff951b1ce6145ca8ecb174/tokenization/data.py#L2012). This filter is designed to remove examples in which the assistant is presented as model other than Lucie (e.g., ChatGPT, Gemma, Llama, ...).

### Instruction template:
Lucie-7B-Instruct-human-data was trained on the chat template from Llama 3.1 with the sole difference that `<|begin_of_text|>` is replaced with `<s>`. The resulting template:

```
<s><|start_header_id|>system<|end_header_id|>

{SYSTEM}<|eot_id|><|start_header_id|>user<|end_header_id|>

{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{OUTPUT}<|eot_id|>
```


An example:


```
<s><|start_header_id|>system<|end_header_id|>

You are a helpful assistant.<|eot_id|><|start_header_id|>user<|end_header_id|>

Give me three tips for staying in shape.<|eot_id|><|start_header_id|>assistant<|end_header_id|>

1. Eat a balanced diet and be sure to include plenty of fruits and vegetables. \n2. Exercise regularly to keep your body active and strong. \n3. Get enough sleep and maintain a consistent sleep schedule.<|eot_id|>
```

### Training procedure

The model architecture and hyperparameters are the same as for [Lucie-7B](https://huggingface.co/OpenLLM-France/Lucie-7B) during the annealing phase with the following exceptions:
* context length: 4096<sup>*</sup> 
* batch size: 1024
* max learning rate: 3e-5
* min learning rate: 3e-6

<sup>*</sup>As noted above, while Lucie-7B-Instruct is trained on sequences of 4096 tokens, it maintains the capacity of the base model, Lucie-7B, to handle context sizes of up to 32K tokens.

## Testing the model

### Test with ollama

* Download and install [Ollama](https://ollama.com/download)
* Download the [GGUF model](https://huggingface.co/OpenLLM-France/Lucie-7B-Instruct-human-data/resolve/main/Lucie-7B-q4_k_m.gguf)
* Copy the [`Modelfile`](Modelfile), adpating if necessary the path to the GGUF file (line starting with `FROM`).
* Run in a shell:
    * `ollama create -f Modelfile Lucie`
    * `ollama run Lucie`
* Once ">>>" appears, type your prompt(s) and press Enter.
* Optionally, restart a conversation by typing "`/clear`"
* End the session by typing "`/bye`".

Useful for debug:
* [How to print input requests and output responses in Ollama server?](https://stackoverflow.com/a/78831840)
* [Documentation on Modelfile](https://github.com/ollama/ollama/blob/main/docs/modelfile.md#parameter)
   * Examples: [Ollama model library](https://github.com/ollama/ollama#model-library)
      * Llama 3 example: https://ollama.com/library/llama3.1
* Add GUI : https://docs.openwebui.com/

### Test with vLLM

#### 1. Run vLLM Docker Container

Use the following command to deploy the model,
replacing `INSERT_YOUR_HF_TOKEN` with your Hugging Face Hub token.

```bash
docker run --runtime nvidia --gpus=all \
    --env "HUGGING_FACE_HUB_TOKEN=INSERT_YOUR_HF_TOKEN" \
    -p 8000:8000 \
    --ipc=host \
    vllm/vllm-openai:latest \
    --model OpenLLM-France/Lucie-7B-Instruct-human-data
```

#### 2. Test using OpenAI Client in Python

To test the deployed model, use the OpenAI Python client as follows:

```python
from openai import OpenAI

# Initialize the client
client = OpenAI(base_url='http://localhost:8000/v1', api_key='empty')

# Define the input content
content = "Hello Lucie"

# Generate a response
chat_response = client.chat.completions.create(
    model="OpenLLM-France/Lucie-7B-Instruct-human-data",
    messages=[
        {"role": "user", "content": content}
    ],
)
print(chat_response.choices[0].message.content)
```

## Citation

When using the Lucie-7B-Instruct-human-data model, please cite the following paper:

✍ Olivier Gouvert, Julie Hunter, Jérôme Louradour,
Evan Dufraisse, Yaya Sy, Pierre-Carl Langlais, Anastasia Stasenko,
Laura Rivière, Christophe Cerisara, Jean-Pierre Lorré (2025)
Lucie-7B LLM and its training dataset
```bibtex
@misc{openllm2023claire,
      title={The Lucie-7B LLM and the Lucie Training Dataset:
      open resources for multilingual language generation}, 
      author={Olivier Gouvert and Julie Hunter and Jérôme Louradour and Evan Dufraisse and Yaya Sy and Pierre-Carl Langlais and Anastasia Stasenko and Laura Rivière and Christophe Cerisara and Jean-Pierre Lorré},
      year={2025},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

## Acknowledgements

This work was performed using HPC resources from GENCI–IDRIS (Grant 2024-GC011015444). We gratefully acknowledge support from GENCI and IDRIS and from Pierre-François Lavallée (IDRIS) and Stephane Requena (GENCI) in particular.

Lucie-7B was created by members of [LINAGORA](https://labs.linagora.com/) and the [OpenLLM-France](https://www.openllm-france.fr/) community, including in alphabetical order:
Olivier Gouvert (LINAGORA),
Ismaïl Harrando (LINAGORA/SciencesPo), 
Julie Hunter (LINAGORA),
Jean-Pierre Lorré (LINAGORA),
Jérôme Louradour (LINAGORA),
Michel-Marie Maudet (LINAGORA), and
Laura Rivière (LINAGORA).


We thank 
Clément Bénesse (Opsci), 
Christophe Cerisara (LORIA),
Evan Dufraisse (CEA),
Guokan Shang (MBZUAI), 
Joël Gombin (Opsci), 
Jordan Ricker (Opsci), 
and
Olivier Ferret (CEA) 
for their helpful input.

Finally, we thank the entire OpenLLM-France community, whose members have helped in diverse ways.

## Contact

contact@openllm-france.fr