Modelpublic

ClassiCC-Corpus/Curio-1.1b

sourceHugging Faceupdated 1y agoView on Hugging Face

3likes34downloads

Model Card

🐦 Curió 1.1B

📖 Overview

Curió 1.1B is a Portuguese-adapted language model created via continued pretraining of TinyLlama 1.1B (1T), originally trained on 1 trillion English tokens, on 150B Portuguese tokens from the ClassiCC-PT corpus.

This model was designed to explore the impact of language-specific corpora on adapting an English-trained base model to Portuguese, yielding performance improvements on Portuguese benchmarks without large-scale retraining from scratch.

🏗 Training Setup

—Base model: TinyLlama 1.1B (LLaMA-2 architecture)

—Parameters: 1.1B

—Continued pretraining tokens: 150B (ClassiCC-PT)

—Sequence length: 4096 tokens (with packing)

—Hardware: TPU v2-128 (thanks to Google TRC program)

—Frameworks: T5X

📊 Evaluation

Evaluated on the Poeta benchmark — 14 diverse Portuguese tasks (RTE, STS, MCQ exams, sentiment analysis, QA, etc.) — using the Normalized Preferred Metric (NPM).

Model	Training Regimen	Poeta v2 NPM
TinyLlama 1T (EN)	–	17.4
TinyLlama 2T (EN)	+1T EN continued pretraining	20.9
training with mC4-PT	+150B PT (mC4-PT) continued pretraining	\~20
training with ClueWeb-22-PT	+150B PT (Clueweb-22-PT) continued pretraining	\~27
Curió 1.1B	+150B PT (ClassiCC-PT) continued pretraining	27.1

📥 Usage

Please note that Curio 1.1B has not trained to be used as a chat model

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "ClassiCC-Corpus/Curio-1.1B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

📜 Citation

If you use Curió 1.1B, please cite:

Coming soon

Acknowledgements

We thank the google TRC program, which generously granted us the necessary resources for the development of this research.