CataLlama-v0.1-Instruct-SFT

CataLlama-v0.1-Instruct-SFT is an instruct fine-tune of catallama/CataLlama-v0.1-Base on the catallama/Catalan-Instruct dataset.

CataLlama was trained on roughly 445 million new tokens in three separate stages. This is the 2nd stage of the training.

The model shows improved proficiency with the Catalan language.

This is an instruction fine-tuned model proficient on the following tasks in Catalan

• Information extraction (suitable for RAG)
• Named Entity Recognition (NER)
• Translation from English to Catalan and Catalan to English
• Summarization - both short form and long form
• Sentiment analysis

The model achieves a loss rate of 0.8528 on the validation dataset after two epochs.

NOTE: The model was trained for one epoch on the train split of dataset and after manual evaluation, I decided to go for another epoch.

The first epoch logs every 100 steps while the second epoch logs every 200 steps, but I am pasting the train and eval losses for both epochs bellow.

The train split of the dataset was shuffled before the second epoch. The test split dataset is identical in both epochs without shuffling

Model developers Laurentiu Petrea based on Llama-3 from Meta.

Model Architecture CataLlama is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and direct preference optimisation (DPO) to align with human preferences for helpfulness and safety.

License The model uses the llama-3 license available at: https://llama.meta.com/llama3/license

Use with transformers

See the snippet below for usage with Transformers:

The model follows the same prompt template as Llama-3 Instruct

import transformers
import torch

model_id = "catallama/CataLlama-v0.1-Instruct-SFT"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Ei com estàs avui?"},
]

prompt = pipeline.tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True
)

outputs = pipeline(
    prompt,
    max_new_tokens=1024,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)

print(outputs[0]["generated_text"][len(prompt):])

Training procedure

The model was trained with the same prompt template of Llama-3 Instruct.

The model was trained for two epochs on 6x A100 80GB GPUs using DeepSpeed ZeRO State-3 without CPU offloading.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• distributed_type: multi-GPU
• num_devices: 6
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 2

Training results

Epoch 1

Training Loss	Epoch	Step	Validation Loss
1.0938	0.11	100	1.0779
1.0186	0.22	200	1.0209
1.0157	0.32	300	0.9808
0.9588	0.43	400	0.9489
0.9039	0.54	500	0.9244
0.9111	0.65	600	0.9086
0.8918	0.75	700	0.8961
0.8971	0.86	800	0.8886
0.8631	0.97	900	0.8846

Epoch 2

Training Loss	Epoch	Step	Validation Loss
0.8002	0.22	200	0.8989
0.8068	0.43	400	0.8835
0.7722	0.65	600	0.8654
0.7805	0.86	800	0.8528

Intended Use

Note: This model is not intended to beat benchmarks, but to demonstrate techniques for augmenting LLMs on new languages and preserve rare languages as part of our world heritage.

Intended Use Cases Llama 3 is intended for commercial and research use in English. Instruction tuned models are intended for assistant-like chat, whereas pretrained models can be adapted for a variety of natural language generation tasks.

Out-of-scope Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by the Acceptable Use Policy and Llama 3 Community License. Use in languages other than English**.

**Note: Developers may fine-tune Llama 3 models for languages beyond English provided they comply with the Llama 3 Community License and the Acceptable Use Policy.