Lumen 1.3 COMMING SOON!
#2
by AxionLabsAI - opened
Lumen 1.3 Balanced retrain:
A full retrain (not a patch) from Qwen3-8B, targeting 70%+ HumanEval and 40%+ GSM8K simultaneously — the dual target 1.2/1.2.1 couldn't hit. Key changes:
- Better data ratio: more reasoning data (MetaMathQA), 3 epochs, lower LR — fixing the 42k-code/15k-math imbalance that caused 1.2.1's tradeoffs
- Safety training data: explicit shutdown-compliance and self-preservation-ethics examples added after 1.2.1's safety eval revealed alignment gaps (14% automated pass rate, 26% manual)
- DPO fine-tuning pipeline on Colab (A100) via trl / peft / llama.cpp — ready to go, just needs the SFT base first
- Timeline extended to get both benchmarks and safety right in one shot
Status: In development.
Oh.. and 1 secret that I can't tell you.
ETA: July 20th, 2026.