README.md

metadata

language:
  - en
  - pt
tags:
  - code
  - instruction-tuning
  - portuguese
  - reasoning
  - cosmos-9b
  - education
size_categories:
  - 100K<n<1M

COSMOS 9B Dataset

Dataset Description

This dataset is a meticulously curated mixture of 100,000 high-quality instruction-response pairs designed specifically for fine-tuning the COSMOS 9B model. It aims to create an AI model that excels as an educator and coding assistant, with strong reasoning capabilities and fluency in Brazilian Portuguese (PT-BR).

The dataset includes injected Chain-of-Thought (...) reasoning steps in Portuguese for a subset of the coding and logic prompts, teaching the model to reason in PT-BR before providing the final answer.

Dataset Composition

• CodeFeedback-Filtered-Instruction (25k): Advanced programming and logic queries.
• Magicoder-Evol-Instruct (20k): Complex, evolved coding trajectories.
• evol-codealpaca-v1 (20k): C/C++, Python, and algorithmic problem solving.
• Magicoder-OSS-Instruct (10k): Grounded, realistic open-source coding instructions.
• Pt-Corpus-Instruct (15k): Deep Brazilian Portuguese conversational fluency.
• SlimOrca (10k): Advanced general reasoning and instruction following (crucial for ENEM/Vestibular logic).

Intended Use

This dataset is specifically intended for the Supervised Fine-Tuning (SFT) of the COSMOS 9B V1 base merge model.