Dev open-sources Peacebell, a from-scratch 291M WWII domain LLM built over 11 months
wayneworkman · reddit · 2026-10-03
- Reddit user wayneworkman spent 11 months of nights and weekends training Peacebell, a from-scratch WWII-specialized small language model, open-sourcing weights and training materials in 291M and 148M versions (the latter for sub-150M HF leaderboards).
- All training data is self-made synthetic data from WWII Wikipedia articles; most time went to data curation. He also built a private WWII benchmark to avoid contamination, plus a custom vLLM fork and a free demo.
- He describes being emotionally devastated reading WWII atrocities during data work; next version expected in 2027.
More from Research
- CogGym Finds Bigger, Newer Models Behave More Like Humans — But Even More Like Each Other — teortaxesTex · 2026-10-03
- PhantomEnvironments: 7B LLM Trained in Synthetic RL Environments Matches Agents 10x Its Size — CShorten30 · 2026-10-03
- Harvard Sinclair lab unveils early mouse data for anti-aging molecule SL-100 — _AustinCalvert_ · 2026-10-03
- MIT Interactive Diagrams: From Attention to Mixtral and DeepSeek-V3 Architectures — vtabbott_ · 2026-10-03
- Deriving KV-cache placement from abstract representations: prefill and inference are linked — vtabbott_ · 2026-10-03
- Luminance dominates geometry formation in 3D Gaussian Splatting, new paper finds — kwangmoo_yi · 2026-10-03