Dev Fine-Tuned a 2B LLM on WhatsApp Group Chat, Simulating Six Friends on an M1 Pro
BarisSayit · reddit · 2026-09-12
A developer spent months fine-tuning a 2B local model on his own six-person WhatsApp group chat, training and running it on an M1 Pro to simulate the whole group. All experiments were in Turkish.
How good is it:
- No coherent group simulation, but it picked up the group's slang, reactions, and pacing — generated messages closely resemble what members would type;
- Some coherence but no deep understanding; the model doesn't retain personal facts about members.
On paper: Using a judge-LLM, human-anchored evaluation, the best version achieved an 80% human win rate in human-vs-model tests (ideal would be <50%).
The author open-sourced the reproducible local pipeline, chat UI, evaluation method, results, and an experiment PDF — but not the private chat data or fine-tuned weights — and reminds you to ask for consent before training on others' chats.
More from Fun
- Dev shows off Freaky Friday, a trippy short made with his new app Basic Slop — bennash · 2026-09-12
- Path tracing and physics sims in a terminal: no_std Rust, fixed-point math, 320x200 256-color — bilawalsidhu · 2026-09-12
- REK Simulator Champions Fly to SF to Pilot Real 6ft Humanoids in Robot Fight — cixliv · 2026-09-12
- Making the worm connectome dance to Midnight City — neuroecology · 2026-09-12
- Gen Z Survey: Healthcare Overtakes 'Influencer' as Top Dream Job in the US — Polymarket · 2026-09-12
- Opus 5 made a choir of faces sing a good morning song, and it's gloriously uncanny — repligate · 2026-09-12