pauhidalgoo/cucafera-instruct

Model Card for cucafera 🔥🐲 (Instruct Model)

This document describes cucafera (Instruct Model), a Catalan Large Language Model (LLM) fine-tuned to follow instructions and generate text in Catalan. Built upon the base model, it leverages high-quality Catalan datasets and is optimized for instruction following tasks.

Model Details

Model Description

cucafera (Instruct Model) is a 244-million parameter transformer-based language model inspired by the LLAMA architecture (notably LLAMA3). Despite its relatively small size compared to many contemporary models, it is optimized for generating coherent and contextually relevant text in Catalan.

• Model Size: 244M parameters
• Architecture: Transformer-based (LLAMA-inspired) with 30 layers
• Embedding Size: 768
• Attention Mechanism: 4 key/value heads and 8 query heads (using Grouped Query Attention - GQA)
• Context Length: 2048 tokens
• Tokenizer: Byte-Pair Encoding (BPE) with a vocabulary size of 65,536
• Activation Function: GeGLU

Instruct Fine-Tuning

The instruct version of cucafera has been fine-tuned on a variety of instruction datasets to enhance its ability to follow user prompts. The fine-tuning was performed using Hugging Face's SFTTrainer and follows the ChatML format for conversation, for example:

<|im_start|>user Fes un poema <|im_end|> <|im_start|>assistant

Training Data

The base model was pre-trained using the patufet-pretrain dataset.

The fine-tuning data utilized a mix of instruction datasets from the patufet collection.

Fine-tunning Procedure

The model was fine-tuned with the following setup:

• Total fine-tunning steps: 1500
• Per device train batch size: 12
• Sequence Length: 2048
• Learning rate: 3e-5
• Optimizer: AdamW
• Weight decay: 0.01
• Epochs: 5

Different commits represent different fine-tunning procedures: we experimented with different data mixes, epochs, datasets...

Direct Use

The cucafera (Instruct Model) is designed for:

• Conversational agents and chatbots in Catalan.
• Task-specific applications such as summarization, translation (within Catalan), and creative writing.
• Educational and experimental research into instruction-following LLMs.
• Creative content generation, like poems or stories

However, due to its limited size, it is not able to provide correct factual information and you must be aware of this fact when using this model.

Out-of-Scope Uses

• High-Stakes Applications:
  The model is not recommended for uses where extremely high factual accuracy is required or where outputs could have significant real-world consequences.
• Non-Catalan Tasks:
  Since the model is exclusively trained on Catalan text, it is not suited for tasks in other languages without further training or fine-tuning.
• Sensitive or safety-critical uses: It has not undergone RLHF/DPO tuning, so outputs should be reviewed carefully.

Bias, Risks, and Limitations

• The model has no instruction tuning, so it may not follow prompts effectively.
• It only understands Catalan, meaning it is unsuitable for multilingual applications.
• Due to its small size (244M parameters), its knowledge and reasoning capabilities are limited.
• It was trained on a limited dataset, which may introduce biases in its outputs.

Recommendations

• The goal of this model is educational. You are encouraged to train your own model.
• If used in production, human review of its outputs is recommended.
• Fine-tuning on task-specific data can improve accuracy and mitigate biases.
• Users should be cautious when using it in sensitive or high-stakes applications.

Use the Instruct Model

You can use the instruct model via huggingface's transformers library. Make sure to specify the ChatML format.

Acknowledgements

This model was developed as an experimental project, inspired by Karpathy's NanoGPT Series. My colleague Roger Baiges also trained his own CatGPT.

For more details, updates, or to contribute to the project, please visit the GitHub repository