This model is a pedagogical instruction fine-tune of Llama-3-8B trained on the Brazilian Base Nacional Comum Curricular (BNCC).
Its goal is to generate clear, observable, and pedagogically valid learning objectives (skills) from competencies and content topics.

Model Details

Model Description

The model was developed as part of a graduate research project at UFRGS (BPM Research Lab). It explores transfer learning from K--12 educational standards (BNCC) to Higher Education curriculum design, with a focus on Business Process Management (BPM).

By learning structured mappings from Competency → Skill present in the BNCC, the model captures what can be described as a pedagogical syntax (verb + object + condition), aligned with Bloom's Taxonomy and instructional design best practices.

Developed by: Guilherme Rockembach\
Funded by: Not applicable (academic research)\
Shared by: Guilherme Rockembach\
Model type: LoRA Adapter (instruction fine-tuned LLM)\
Language(s) (NLP): Portuguese (pt-BR)\
License: Llama 3 Community License\
Finetuned from model: unsloth/llama-3-8b-bnb-4bit

Model Sources

Repository: https://huggingface.co/guirrock/llama3-8b-bncc-pedagogy\
Base framework: https://github.com/unslothai/unsloth

Uses

Direct Use

The model is intended for:

Instructional designers\
Teachers and curriculum developers\
Educational researchers

It generates learning skills/objectives from a given educational area and competency, following pedagogical taxonomies and BNCC-style formulations.

Downstream Use

The model can be integrated into automated curriculum design frameworks, where it serves as the generative component for learning objectives derived from structured educational data or mined curricular documents.

Out-of-Scope Use

General-purpose chat or conversational agents\
Mathematical reasoning or symbolic computation\
Code generation\
Non-educational or non-curricular tasks

Bias, Risks, and Limitations

The model reflects the normative structure and assumptions of the BNCC, which may not generalize to all educational systems.
Outputs may overemphasize Bloom-style formulations even when alternative pedagogical approaches are desired.
The model does not validate pedagogical correctness against real classroom outcomes.

Recommendations

Users should treat generated skills as drafts or suggestions, subject to expert review and contextual adaptation. The model should support---not replace---human pedagogical judgment.

How to Get Started with the Model

The model can be efficiently used with the Unsloth library for fast and memory-efficient inference.

python

from unsloth import FastLanguageModel
import torch

model_name = "guirrock/llama3-8b-bncc-pedagogy"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name,
    max_seq_length=2048,
    load_in_4bit=True,
)

FastLanguageModel.for_inference(model)

Training Details

Training Data

Source: Brazilian Ministry of Education (MEC) -- BNCC\
Structure: Area / Competency → Skill\
Domains: Mathematics, Languages, Natural Sciences, Human Sciences\
Size: Approximately 1k--5k examples

Training Procedure

Framework: Unsloth (QLoRA)
Quantization: 4-bit (bnb-4bit)

Training Hyperparameters

LoRA rank: 16\
LoRA alpha: 16\
Optimizer: AdamW (8-bit)\
Training regime: 4-bit quantized fine-tuning

Speeds, Sizes, Times

GPU: NVIDIA Tesla T4\
Environment: Google Colab (Free Tier)\
Training time: < 1 hour

Evaluation

Testing Data, Factors & Metrics

Testing data: Held-out BNCC-style competency--skill pairs\
Factors: Educational domain and competency type\
Metrics: Qualitative pedagogical analysis

Results

The model demonstrates strong transfer of pedagogical structure from K--12 standards to higher-education BPM contexts, producing skills aligned with Bloom's Taxonomy without relying on rigid rule-based templates.

Environmental Impact

Hardware Type: NVIDIA Tesla T4\
Hours used: ~0.5\
Cloud Provider: Google Colab\
Compute Region: Not specified\
Carbon Emitted: < 0.1 kg CO₂eq (estimated)

Technical Specifications

Model Architecture and Objective

Decoder-only transformer (Llama 3)
Objective: Instruction-following generation of pedagogical learning objectives

Compute Infrastructure

Hardware

NVIDIA Tesla T4 GPU

Software

PEFT 0.18.1\
Unsloth\
Transformers

Citation

BibTeX:

bibtex

@mastersthesis{rockembach2026bpm,
  author  = {Rockembach, Guilherme},
  title   = {Framework for Automated Competency-Based Curriculum Design in BPM},
  school  = {Universidade Federal do Rio Grande do Sul (UFRGS)},
  year    = {2026}
}

Model Card Authors

Guilherme Rockembach

Model Card Contact

Guilherme Rockembach

Framework Versions

PEFT 0.18.1