Logo Lanfrica

Mystique1337/telcoproject

Domain:

natural language processing

Record type:

modelsoftware
Creator:
Mys
Host:
Naija Persona Agent: Nigerian-context AI for review simulation (InsideNaija) and product recommendation (ShopEasy), powered by NaijaReviewer-8B, a fine-tuned Llama 3.1 8B. # Naija Persona Agent **A Nigerian-context LLM agent system for review simulation and personalised product recommendation.** Submission to the **DSN X BCT LLM Agent Challenge**. - **Live app:** - **Paper, Task A (User Modelling):** `paper/USER_MODELLING_TASK.pdf` - **Paper, Task B (Recommendation):** `paper/RECOMMENDATION_PAPER.pdf` - **Model weights (GGUF):** - **Training corpus:** --- ## The thesis General-purpose LLM agents trained on predominantly Western web data carry an implicit **cultural prior**. When prompted to simulate a Nigerian shopper, frontier models compress rating intensity, flatten Pidgin and Nigerian-English register into a sanitised international voice, default to individualistic framing, and routinely misread cultural and religious markers. This project measures that gap and recovers it for the Nigerian context through three coordinated mechanisms: 1. A structured **cognitive persona** representation (four behavioural dimensions plus a Nigerian register tier). 2. Register-aware prompting grounded in that persona. 3. **NaijaReviewer-8B**, an open-weight Llama 3.1 8B model QLoRA-fine-tuned on Nigerian product reviews, served as the review-generation and re-ranking backbone. The system ships as a FastAPI service plus a React frontend that exposes two products. ## Headline results **Task A (review generation), 100-row held-out test split.** | Metric | NaijaReviewer-8B | Claude Sonnet 4 | |---|---|---| | RMSE (lower is better) | **1.114** | 1.319 | | BERTScore F1 | 0.858 | 0.857 | | AgentSociety `preference_estimation` | **0.792** | 0.784 | | Human-rater win-rate (5 raters, 50 pairs) | **48.5%** (CI 40.2 to 56.9) | 51.5% | A 15.5% RMSE reduction over Claude at roughly 17 times fewer parameters; Nigerian human raters judge the two systems at parity on prose authenticity. **Task B (recommendation), held-out persona scenarios.** | Re-ranker | Params | NDCG@10 | HR@5 | |---|---|---|---| | **NaijaReviewer-8B** | **8B** | **0.588** | **0.648 …