Historical Event Narrator
LoRA fine-tuned Mistral-7B system for creative alternate-history generation and live LLM-as-judge evaluation.
Project information
- Category: Fine-Tuned LLM / GenAI
- Dates: Oct 2025 - Dec 2025
- Tech: Python, PyTorch, Hugging Face Transformers, PEFT, LoRA, Mistral-7B, Gemini API, Streamlit, Apple Silicon MPS
- GitHub: Historical-Event-Narrator
Project Details
Overview
Historical Event Narrator fine-tunes a Mistral-7B based model to generate creative alternate-history narratives by combining factual historical events with plausible what-if twists.
Problem
Base open-source models were factual but dry, producing essay-like summaries with weak creative divergence. The goal was to teach a smaller local model to generate vivid, cinematic narratives while preserving historical relevance.
Engineering
- Fine-tuned Mistral-7B-v0.1 with LoRA and PEFT, using parameter-efficient updates instead of full retraining.
- Trained locally on Apple Silicon MPS with mixed precision to work within hardware constraints and avoid paid cloud GPUs.
- Generated a synthetic dataset by transforming 1,200+ Wikipedia articles into narrative-plus-twist pairs with Gemini 2.5 Flash.
- Implemented guided inference, stopping criteria, and a Streamlit evaluation layer using Gemini Pro as an LLM judge for creativity, relevance, and twist quality.
Impact
A/B evaluation moved outputs from factual summaries to vivid alternate narratives, improving creativity from 3/10 to 8/10, twist quality from 0/10 to 9/10, and ROUGE-L structural coherence by 100%+ for the task.