FLOR-6.3B is a 6.3B-parameter transformer-based causal language model for Catalan, Spanish, and English. It is the result of a language adaptation technique performed on BLOOM-7.1B, which involves modifying the model's vocabulary and embedding layer, and continuously pre-training the model with 140B tokens in our target languages.

For more details, take a look at this blogpost about the project.

The FLOR-6.3B model is ready-to-use only for causal language modeling. It can perform text-generation tasks and be fine-tuned for specific scenarios.

import torch
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

input_text = "Sovint em trobo pensant en tot allò que"

model_id  = "projecte-aina/FLOR-6.3B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
generator = pipeline(
    "text-generation",
    model=model_id,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)
generation = generator(
    input_text,
    do_sample=True,
    top_k=10,
    eos_token_id=tokenizer.eos_token_id,
)

print(f"Result: {generation[0]['generated_text']}")

At the time of submission, no measures have been taken to estimate the bias and toxicity embedded in the model. However, we are well aware that our models may be biased since the corpora have been collected using crawling techniques on multiple web sources. We intend to conduct research in these areas in the future, and if completed, this model card will be updated.

The language adaptation technique used to create FLOR-6.3B requires the vocabulary of the source model to be adapted before continuing its pre-training with data in the target languages. Specifically, we proceeded as follows:

1. We trained our own BPE tokenizer for Catalan, Spanish, and English, and replaced the original BLOOM tokenizer and vocabulary with it. This procedure implied a downsizing of the original BLOOM's embedding layer and, therefore, a model compression from 7.1B parameters to 6.3B.
2. The embeddings corresponding to tokens that are present in both the original and the target vocabulary (matching tokens) were used for initialization.
3. The embeddings from tokens not present in BLOOM's original vocabulary were initialized as the average of all embeddings.
4. The model was initialized with the weights from BLOOM-7.1B, and with our adapted tokenizer (step 1) and embeddings (steps 2-3).
5. The model was then trained on a corpus that contains a mixture of Catalan, Spanish, and English data.

The training corpus is composed of 140B tokens gathered from web crawlings and public domain data. Most of the sources in Catalan have been obtained from the CATalog 1.0 dataset, filtered with a minimum threshold of 0.6 and oversampling some of the sources it integrates to different extents.

The training data has the same amount of Catalan, Spanish, and English texts. The table below shows the final language distribution:

The training was conducted in 16 Cerebras' CS-2 systems using the cs-2.0.2 release of their software.

FLOR-6.3B has been evaluated in a 5-shot setting, using EleutherAI's LM Evaluation Harness. The evaluation benchmark includes tasks in Catalan, Spanish, and English, with particular emphasis on Catalan datasets.

The tasks were chosen to cover several evaluation areas in order to provide a comprehensive overview of the model's capabilities. The baselines used to compare our results are multilingual and English open-source 7B models and smaller models of the FLOR family of models: TBC.

Our implementation of EleutherAI's LM Evaluation Harness can be found here.

The following is a list of evaluation areas and their respective datasets:

• Reading Comprehension: Belebele
• Question Answering: XQuAD, CatalanQA, CoQCat
• Natural Language Inference: XNLI and its translation to Catalan (XNLI-ca), TE-ca
• Paraphrase Identification: PAWS-X and its translation to Catalan (PAWS-ca), Parafraseja
• Commonsense Reasoning: COPA and its translation to Catalan (COPA-ca)
• Translation: Flores-200

Note: The metrics are F1-score for question-answering tasks, BLEU for translation, and accuracy for the rest.

The Language Technologies Unit from Barcelona Supercomputing Center.

For further information, please send an email to [email protected].

Copyright(c) 2023 by Language Technologies Unit, Barcelona Supercomputing Center.

Apache License, Version 2.0

This work was funded by [Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya](https://politiquesdigitals.gencat.cat/ca/inici/index.html#googtrans(ca|en) within the framework of Projecte AINA.

{
  "input_data": {
    "input_string": [
      "Once upon a time,"
    ],
    "parameters": {
      "top_p": 0.8,
      "temperature": 0.8,
      "max_new_tokens": 90,
      "do_sample": true
    }
  }
}

[
  {
    "0": "Once upon a time, there was a village where the villagers lived in peace and harmony. They worked together, shared their food and resources, and lived in a way that made them happy.\n\nOne day, a stranger arrived in the village. He was a wise and powerful man who could see the future. He told the villagers that their way of life was not sustainable and that they needed to change it.\n\nThe villa"
  }
]

Version: 1

SharedComputeCapacityEnabled hiddenlayerscanned huggingface_model_id : projecte-aina/FLOR-6.3B inference_compute_allow_list : ['Standard_NC6s_v3', 'Standard_NC12s_v3', 'Standard_NC24s_v3', 'Standard_NC24ads_A100_v4', 'Standard_NC48ads_A100_v4', 'Standard_NC96ads_A100_v4', 'Standard_ND96asr_v4', 'Standard_ND96amsr_A100_v4'] inference_supported_envs : ['vllm'] license : apache-2.0 disable-batch : true task : text-generation

View in Studio: https://ml.azure.com/registries/azureml/models/projecte-aina-FLOR-6-3B/version/1

License: apache-2.0

SharedComputeCapacityEnabled: True

SHA: ed6840476f5e9dc1ec693406ab7a45e8acfc2aca

datasets:

inference-min-sku-spec: 6|1|112|64

inference-recommended-sku: Standard_NC6s_v3, Standard_NC12s_v3, Standard_NC24s_v3, Standard_NC24ads_A100_v4, Standard_NC48ads_A100_v4, Standard_NC96ads_A100_v4, Standard_ND96asr_v4, Standard_ND96amsr_A100_v4

languages: en, es, ca