Downloads · 30 days
16
6% of all-time downloads
Shinzmann/naija-reviewer-8b-v2-Q5_0-GGUF
naija-reviewer-8b-v2-Q5_0-GGUF is a text generation model from Shinzmann. Use it when you need the model to write or continue text. It is set up for gguf. The card lists the license as llama3.1.
Llama 3.1 8B Instruct QLoRA fine-tuned on Nigerian product reviews. This repository is the Q50 GGUF build, the higher-quality quantisation in the NaijaReviewer-8B family: roughly 5 bits per weight, ~6 GB on disk, near…
Downloads · 30 days
16
6% of all-time downloads
All-time downloads
285
Public
Repo size
5.6 GB
Likes
0
Public
Click a slice to open those files.
.gguf5.6 GB · 100%
From the Hugging Face model README
Llama 3.1 8B Instruct QLoRA fine-tuned on Nigerian product reviews. This repository is the Q5_0 GGUF build, the higher-quality quantisation in the NaijaReviewer-8B family: roughly 5 bits per weight, ~6 GB on disk, near-imperceptible quality drop versus FP16 while still running comfortably on a single consumer or commodity-cloud GPU.
NaijaReviewer-8B 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.
| Build | Bits | Size | Quality | Use case |
|---|---|---|---|---|
| Q4_K_M | ~4 | ~5 GB | Balanced (Ollama default) | Production serverless, smaller VMs, mobile-class inference. |
| Q5_0 (this repo) | ~5 | ~6 GB | Recommended for quality | Local inference where quality matters more than disk footprint. |
| Q8_0 | 8 | ~8.5 GB | Near-lossless | Reference / sanity-check runs. |
Choose Q5_0 if you have ~6 GB of free VRAM (or system RAM) and want a quality-leaning quant.
| Metric | NaijaReviewer-8B | Frontier baseline (Claude Sonnet 4) |
|---|---|---|
| Task A rating RMSE (lower is better) | 1.114 | 1.319 (15.5% higher) |
| 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 four heavyweight baselines | 0.588 (best in field) | 0.430-0.441 |
| Parameters | 8B | 70B-120B+ |
| Per-call API cost | Zero (open weights) | $/1k tokens |
naija-reviewer-8b-v2-Q5_0.gguf — the quantised model (~6 GB).Modelfile — Ollama configuration encoding the exact Alpaca prompt template the model was trained on, with the correct stop tokens.huggingface-cli download Shinzmann/naija-reviewer-8b-v2-Q5_0-GGUF \
naija-reviewer-8b-v2-Q5_0.gguf Modelfile --local-dir .
ollama create naija-reviewer-8b -f Modelfile
ollama run naija-reviewer-8b "Write a review of a Tecno Spark 10 phone from a Lagos Bolt driver."
huggingface-cli download Shinzmann/naija-reviewer-8b-v2-Q5_0-GGUF \
naija-reviewer-8b-v2-Q5_0.gguf --local-dir .
./llama-cli -m naija-reviewer-8b-v2-Q5_0.gguf \
-p "### Instruction\nWrite a product review.\n\n### Input\nA Tecno Spark 10 mobile phone, NGN 145,000.\n\n### Response\n" \
-n 256 --temp 0.7 --top-p 0.9 -no-cnv
from llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="Shinzmann/naija-reviewer-8b-v2-Q5_0-GGUF",
filename="naija-reviewer-8b-v2-Q5_0.gguf",
n_ctx=4096,
n_gpu_layers=-1,
chat_format="llama-3",
)
out = llm.create_chat_completion(
messages=[{
"role": "user",
"content": "Write a short, authentic Pidgin review of an Oraimo wireless earbud."
}],
max_tokens=256,
temperature=0.7,
)
print(out["choices"][0]["message"]["content"])
The reference production deployment uses Q4_K_M for the lowest cold-start cost on a serverless L4 GPU. To swap to Q5_0, change HF_REPO in deploy/modal_naija.py to Shinzmann/naija-reviewer-8b-v2-Q5_0-GGUF and redeploy. Expect ~20 % more memory and slightly slower warm latency in exchange for higher fidelity.
Fine-tuned on the Alpaca template:
### Instruction
{instruction}
### Input
{input}
### Response
{response}
Stop sequences: ### Instruction, ### Input, ### Response. The included Ollama Modelfile encodes these along with a Nigerian-context system prompt.
For Nigerian-context use, pass the structured persona JSON (cognitive dimensions + register tier + aspect priorities) as ### Input. The production prompt templates are in the project repo under app/prompts/.
| 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 via train_on_responses_only |
| Tokenisation | EOS-terminated training examples |
| Schedule | 2 epochs, effective batch size 16, learning rate 2e-4 with cosine decay, sequence length 4096 |
| Quantisation (this file) | Q5_0, ~6 GB |
Trained on Shinzmann/npa-corpus-v1 (~20,000 Alpaca-style instruction/response pairs), built from two real public Jumia sources plus synthetic expansion. Full provenance and the build pipeline are in the dataset card.
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 disjoint train/eval generator families were used to mitigate confounds. 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.
@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-Q5_0-GGUF}
}
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.