metadata

language:
  - en
  - es
  - ca
licence: apache-2.0
tags:
  - spanish
  - catalan
  - falcon-7b
datasets:
  - BSC-LT/open_data_26B_tokens_balanced_es_ca
metrics:
  - ppl
model-index:
  - name: falcon_7b_balanced_tokenizer_fp16_CPT_open_data_26B_tokens_balanced_es_ca
    results:
      - task:
          name: Causal Language Modeling
          type: text-generation
        dataset:
          name: BSC-LT/open_data_26B_tokens_balanced_es_ca
          type: Causal Language Modeling
          config: default
          split: validation
          args: default
        metrics:
          - name: Perplexity
            type: ppl
            value: 8.59
widget:
  - text: |-
      Respòn a la pregunta següent.
      Pregunta: "Qui viu a França?"
      Resposta: "A França viuen els francesos."
      ----
      Respòn a la pregunta següent.
      Pregunta: "Quina és la capital de Suècia?"
      Resposta: "La capital de Suècia és Estocolm."
      ----
      Respòn a la pregunta següent.
      Pregunta: "Quina beguda es consumeix als matins per despertar-se?"
      Resposta: "La majoria de gent consumeix cafè per despertar-se."
      ----
      Respòn a la pregunta següent.
      Pregunta: "Qui és Leo Messi?"
      Resposta:
    example_title: Pregunta-Resposta
  - text: >-
      Extrae las entidades nombradas del siguiente texto:

      Texto: "Me llamo Wolfgang y vivo en Berlin"

      Entidades: Wolfgang:PER, Berlin:LOC

      ----

      Extrae las entidades nombradas del siguiente texto:

      Texto: "Hoy voy a visitar el parc güell tras salir del barcelona
      supercomputing center"

      Entidades: parc güell:LOC, barcelona supercomputing center:LOC

      ----

      Extrae las entidades nombradas del siguiente texto:

      Texto: "Maria y Miguel no tienen ningún problema contigo"

      Entidades: Maria:PER, Miguel:PER

      ----

      Extrae las entidades nombradas del siguiente texto:

      Texto: "Damián se cortó el pelo"

      Entidades: Damián:PER

      ----

      Extrae las entidades nombradas del siguiente texto:

      Texto: "Lo mejor de Barcelona és el bar de mi amigo Pablo"

      Entidades: Pablo:PER, Barcelona:LOC

      ----

      Extrae las entidades nombradas del siguiente texto:

      Texto: "Carlos comparte piso con Marc"

      Entidades:
    example_title: Entidades-Nombradas
license: apache-2.0
pipeline_tag: text-generation

falcon_7b_balanced_tokenizer_fp16_CPT_open_data_26B_tokens_balanced_es_ca

Model description

The Cǒndor-7B is a transformer-based causal language model for Catalan, Spanish, and English. It is based on the Falcon-7B model and has been trained on a 26B token trilugual corpus collected from publicly available corpora and crawlers.

Intended uses & limitations

The Cǒndor-7B model is ready-to-use only for causal language modeling to perform text-generation tasks. However, it is intended to be fine-tuned on a generative downstream task.

How to use

Here is how to use this model:

import torch
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM

input_text = "Maria y Miguel no tienen ningún "
model = "BSC-LT/condor-7b"
tokenizer = AutoTokenizer.from_pretrained(model)

pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)
generation = pipeline(
    input_text,
    max_length=200,
    do_sample=True,
    top_k=10,
    eos_token_id=tokenizer.eos_token_id,
)

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

Limitations and biases

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.

Language adaptation

We adapted the original Falcon-7B model to Spanish and Catalan by swapping the tokenizer and adjusting the embedding layer. The adaptation procedure is explained in this blog.

New vocabulary

We trained a new BPE Tokenizer for the Catalan and Spanish languages (equal representation). We shuffled a small amount of English in the mixture (since English is in the model training data). The resulting data has the following language distribution:

Language	%
En	16.84%
Es	41.38%
Ca	41.79%

This reduced drastically the number of tokens required to tokenize a text in the target language while the English tokenization shows a small increase.

Embedding Layer Initialization

In order to fully take advantage of the English Pre-Training of the original Falcon model, we decided to re-use the embedding weights of the original model for those tokens shared between the two Tokenizers (the new and the old one). The rest of the embedding weights are initialized as the mean value of the weights of the original Tokenizer.

Training

Training data

The training corpus consists 26B tokens of several corpora gathered from web crawlings and public corpora.

Dataset	Language	Tokens (pre-epoch)	Epochs
Wikipedia	en	2169.97M	1.428144485
C4_es	es	53709.80M	0.1049686196
Biomedical	es	455.03M	0.7140722425
Legal	es	995.70M	0.7140722425
Wikipedia	es	693.60M	1.428144485
Gutenberg	es	53.18M	0.7140722425
C4_ca	ca	2826.00M	2.142216727
Biomedical	ca	11.80M	1.428144485
RacoCatalá Noticias	ca	17.16M	2.142216727
RacoCatalá Forums	ca	333.73M	2.142216727
CaWaC	ca	57.79M	2.142216727
Wikipedia	ca	228.01M	3.570361212
Vilaweb	ca	50.34M	2.142216727

The dataset has the following language distribution:

Language	%
En	16.84%
Es	41.38%
Ca	41.79%

Training procedure

The training corpus has been tokenized using a byte version of Byte-Pair Encoding (BPE) used in the original RoBERTA model with a vocabulary size of 50,262 tokens. Once the model has been successfully initialized, we continued its pre-training in the three target languages: Catalan, Spanish, and English. We kept a small amount of English in order to avoid catastrophic forgetting. The training lasted a total of 96 hours with 8 NVIDIA H100 GPUs of 80GB of RAM.