gemma4-12b-sorrel-selfloop-g2-selfjudge-chat

Private research artifact — do not redistribute. Anthropic Fellows project on flourishing-framed character training (Wang & Jermyn pitch, 2026-04-22).

• base model: joshycodes/gemma4-12b-sorrel-selfloop-g2-midtrain @ bd88d4bf5e1f
• steps: chat
• run: gemma4-12b-sorrel-selfloop-g2-midtrain-local-self-5k.jsonl-c-0916-1951 on 4x NVIDIA H200 (RunPod, fellows worker)
• launcher commit: a0afb77669ae of the flourishing-training repo
• seed: 20260821

step	data	revision	tokens seen	loss
chat	local:self-5k.jsonl (config self-5k)	main	3,203,033 tokens	loss 0.9148 → 1.0362

Hyperparameters:

{
  "chat": {
    "lr": 1e-05,
    "seq_len": 4096,
    "micro_batch": 1,
    "grad_accum": 32,
    "epochs": 1.0
  }
}

Full run config in train_run_config.json; evaluate with uv run eval.py --model joshycodes/gemma4-12b-sorrel-selfloop-g2-selfjudge-chat --eval all.