Llama 3.1 8B Instruct QLoRA fine-tuned on Nigerian product reviews. This repository contains both the merged HF-format weights (for transformers / Unsloth) and all GGUF quantisations (for llama.cpp / Ollama). It is the model behind the Naija Persona Agent — a Nigerian-context AI system for review simulation (Task A) and persona-aware recommendation (Task B), submitted to the DSN X BCT LLM Agent Challenge.

Live application: https://switteefranca2-0--naijapersona-web.modal.run/
Source code: https://github.com/Mystique1337/telcoproject
Companion artifacts: GGUF-only repo · LoRA adapter · Training corpus

Headline numbers

Metric	NaijaReviewer-8B	Frontier baseline (Claude Sonnet 4)
Task A rating RMSE (lower is better)	1.114	1.319 (15.5% higher)
Task A BERTScore F1	0.858	0.857
Task A Nigerian-rater win-rate, 5 raters / 50 pairs	48.5% (CI [40.2, 56.9])	51.5% (statistical parity)
Task B NDCG@10 vs 4 baselines	0.588 (best in field)	0.430
Parameters	8B	70B-120B+
Per-call API cost	Zero (open weights)	$/1k tokens

What's in this repo

File set	Use
`model.safetensors`, `config.json`, tokenizer files	HF format. Load with `transformers` / Unsloth / vLLM / TGI.
`naija-reviewer-8b-v2-Q4_K_M.gguf`	Recommended GGUF (~5 GB). Load with `llama.cpp` / Ollama.
`naija-reviewer-8b-v2-Q5_K_M.gguf`	Higher-quality GGUF (~5.7 GB).
`naija-reviewer-8b-v2-Q8_0.gguf`	Near-lossless GGUF (~8.5 GB).
`Modelfile`	Ollama configuration with the correct Alpaca template and stop sequences.

The Q4_K_M GGUF is also available standalone in its dedicated repo, which is the file the production serverless endpoint pulls.

Quick start

transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("Shinzmann/naija-reviewer-8b-v2")
model = AutoModelForCausalLM.from_pretrained(
    "Shinzmann/naija-reviewer-8b-v2",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = (
    "### Instruction\n"
    "Write a star rating and a short product review.\n\n"
    "### Input\n"
    "Persona: Tunde, 28, Lagos Bolt driver, code-mixed speaker, value-priority.\n"
    "Product: Tecno Spark 10 mobile phone, 4 GB RAM, 5000 mAh battery, NGN 145,000.\n\n"
    "### Response\n"
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=220, eos_token_id=tok.eos_token_id, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

llama-cpp-python (GGUF)

python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="Shinzmann/naija-reviewer-8b-v2",
    filename="naija-reviewer-8b-v2-Q4_K_M.gguf",
    n_ctx=4096,
    n_gpu_layers=-1,
    chat_format="llama-3",
)
out = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Write a Pidgin review of an Oraimo wireless earbud."}],
    max_tokens=256, temperature=0.7,
)
print(out["choices"][0]["message"]["content"])

Ollama

bash

huggingface-cli download Shinzmann/naija-reviewer-8b-v2 \
  naija-reviewer-8b-v2-Q4_K_M.gguf Modelfile --local-dir .
ollama create naija-reviewer-8b -f Modelfile
ollama run naija-reviewer-8b

Prompt format

Alpaca template, as used in training:

### Instruction
{instruction}

### Input
{input}

### Response
{response}

Stop sequences: ### Instruction, ### Input, ### Response. The included Ollama Modelfile encodes these.

For Nigerian-context use, pass the structured persona JSON (cognitive dimensions + register tier + aspect priorities) as the ### Input. The production prompt templates are in the project repo at app/prompts/.

Training


Base model	`meta-llama/Meta-Llama-3.1-8B-Instruct`
Method	QLoRA via Unsloth
Adapter	LoRA r=16, alpha=32, dropout=0.1, targets q/k/v/o/up/gate/down (0.52% trainable params)
Loss	Response-only loss (gradients on the answer tokens only)
Tokenisation	EOS-terminated training examples
Schedule	2 epochs, effective batch size 16, learning rate 2e-4 with cosine decay, sequence length 4096
Quantisation	Three GGUF builds: `Q4_K_M`, `Q5_K_M`, `Q8_0`

Training and validation loss curves are published in the project's notebooks/03_training_results.ipynb and reproduced in the Task A solution paper.

Training data

Trained on Shinzmann/npa-corpus-v1 (~20,000 Alpaca-style instruction/response pairs), built from two real public sources plus synthetic expansion:

Real Jumia reviews (~76k) from the aymane-maghouti/Sentiment-Analysis-for-Jumia-Reviews public GitHub repository, used as Pipeline 1 seed.
Real Jumia product catalogue (~18k) from the Idowenst/jumia_dataset HuggingFace dataset, used as Pipeline 2 seed.
Synthetic expansion by NVIDIA Nemotron and OpenAI models, persona-conditioned and stratified 90/5/5 by register tier.

Intended use and limitations

Intended use. Generation of Nigerian-context product reviews and ratings (Task A); persona-aware re-ranking of product recommendations (Task B); research on register-aware text generation in low-resource African contexts.

Limitations. Trained primarily on Nigerian English and Nigerian Pidgin product reviews. The training corpus is partly synthetic; two independent generator pipelines and a held-out evaluation generated by a different model family were used to mitigate confounds, but residual bias remains possible. On a 3-arbiter LLM-as-Judge evaluation, frontier LLM judges showed a systematic preference for Claude Sonnet 4's prose register; Nigerian human raters scored the two systems at parity on the same pairs, which we read as evidence that single-judge LLM evaluation is insufficient on culturally-localised content.

Citation

bibtex

@misc{naijareviewer8b2026,
  title  = {NaijaReviewer-8B: A Nigerian-Context Open-Weight Fine-Tune for Persona-Aware Review Generation},
  author = {Ashinze, Emmanuel and Uvere, Franca and Oyenekan, Esther},
  year   = {2026},
  url    = {https://huggingface.co/Shinzmann/naija-reviewer-8b-v2}
}

License

Subject to the Llama 3.1 Community License. Released for research and non-commercial use; commercial use must comply with the upstream Llama 3.1 terms.