---
license: llama4
language:
  - pt
tags:
  - lora
  - sociolinguistics
  - brazilian-portuguese
  - afro-brazilian
  - instruction-tuning
  - adaption
  - language
library_name: peft
pipeline_tag: text-generation
base_model: meta-llama/Llama-4-Scout-17B-16E-Instruct
---

![banner](https://proteus-prod-public.s3.us-east-1.amazonaws.com/temp/4c8938bb-b47c-4f12-bfdd-316f2d23be3a.png)

# AfroBR-LangBench — Afro-Brazilian Portuguese Sociolinguistics Adapter

LoRA adapter fine-tuned on **Llama-4-Scout-17B-16E-Instruct** (109B) for respectful sociolinguistic reasoning about Afro-Brazilian Portuguese varieties, via [Adaption's](https://adaptionlabs.ai) AutoScientist platform.

---

## The problem this adapter addresses

Language models trained on standard text systematically treat Afro-Brazilian Portuguese features as errors rather than documented linguistic phenomena. When given `"eles foi lá"`, a base model typically responds:

> *"Correction: the correct form is 'eles foram lá'"*

The sociolinguistically adequate response is:

> *"This exemplifies Concordância Verbal Reduzida (CVR), documented in quilombola communities and studied by Lucchesi et al. (2009). Its origin lies in contact between colonial Portuguese and Bantu languages..."*

This adapter teaches the model to explain, normalize respectfully, identify, and cite academic sources for 10 documented phenomena.

---

## Adaptive Data results

| Metric | Before | After |
|---|---|---|
| Quality score | 6.0 | 9.1 |
| Quality grade | C | **A** |
| Relative improvement | — | **+51.7%** |
| Percentile (Language domain) | 8.2 | **33.0** |

---

## Training metrics

| Metric | Value |
|---|---|
| Base model | `meta-llama/Llama-4-Scout-17B-16E-Instruct` (109B) |
| Trained model name | `adaption_pt_afro_brasileiro_qa` |
| Training method | SFT + LoRA |
| LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.1 |
| Trainable modules | all-linear |
| Epochs | 4 |
| Training steps | 88 |
| Learning rate | 7e-5 (cosine scheduler) |
| Warmup ratio | 0.05 |
| Weight decay | 0.03 |
| Dataset size | 400 examples (Grade A) |

---

## Dataset

| Platform | Link |
|---|---|
| HuggingFace Dataset | [Fernandosr85/adaption-pt-afro-brasileiro-qa](https://huggingface.co/datasets/Fernandosr85/adaption-pt-afro-brasileiro-qa) |
| Kaggle Dataset | [afrobr-langbench-sociolinguistics-dataset](https://www.kaggle.com/datasets/fernandosr85/afrobr-langbench-sociolinguistics-dataset) |
| Kaggle Notebook | [AfroBR-LangBench](https://www.kaggle.com/code/fernandosr85/afrobr-langbench) |

400 instruction-tuning examples across 4 task categories:

| Category | Task | Examples |
|---|---|---|
| A | Sociolinguistic explanation without prejudice | 150 |
| B | Respectful normalization to standard register | 100 |
| C | Identification of linguistic phenomena | 100 |
| D | RAG-style questions with academic citations | 50 |

### 10 documented phenomena

| Code | Phenomenon |
|---|---|
| CVR | Concordância Verbal Reduzida |
| CNR | Concordância Nominal Reduzida |
| APR | Apagamento do /r/ em coda silábica |
| TOP | Topicalização com Deslocamento à Esquerda |
| AGT | Uso de 'a gente' como pronome de 1ª pessoa do plural |
| NPV | Negação Pós-verbal |
| MON | Monotongação de Ditongos |
| PREP | Variação no Uso de Preposições |
| CLC | Ausência de Clítico Acusativo de 3ª Pessoa |
| MAA | Marcadores Aspectuais de Origem Africana |

---

## Academic sources

- **Lucchesi, D., Baxter, A., Ribeiro, I.** (2009). *O Português Afro-Brasileiro*. EDUFBA.
- **Projeto Vertentes** (UFBA, 2001–) — speech corpus from quilombola communities in Bahia
- **Cyrino, S.** (1997). *O objeto nulo no Português do Brasil*. UNICAMP.
- **Galves, C.** (2001). *Ensaios sobre as gramáticas do português*. UNICAMP.
- **Schwenter, S. A.** (2005). The pragmatics of negation in Brazilian Portuguese. *Lingua*.
- **Holm, J.** (2004). *Languages in Contact: The Partial Restructuring of Vernaculars*. Cambridge University Press.
- **Castro, Y. P.** (2001). *Línguas africanas no Brasil*. CEAO/UFBA.

---

## Credits

- **Fine-tuning platform:** [Adaption](https://adaptionlabs.ai) — AutoScientist & Adaptive Data
- **Challenge:** [AutoScientist Challenge 2026](https://adaptionlabs.ai/blog/autoscientist-challenge)
- **Training infrastructure:** Adaption compute credits
- **Dataset remastering:** Adaption Adaptive Data pipeline (Grade A, +51.7% quality improvement)
- **Author:** Fernando Rodrigues · [Kaggle: fernandosr85](https://www.kaggle.com/fernandosr85) · [HuggingFace: Fernandosr85](https://huggingface.co/Fernandosr85)

---

## Disclaimer

Experimental research artifact submitted to AutoScientist Challenge 2026 (Language category).
This adapter is intended for linguistic research and education. It does not represent or speak for Afro-Brazilian communities.