qwen3-4b-sorrel-selfloop-g2-midtrain

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

• base model: Qwen/Qwen3-4B-Base @ 906bfd4b4dc7
• steps: midtrain
• run: qwen3-4b-base-sorrel-selfloop-b-g1-m-0915-2359 on 2x NVIDIA H200 (RunPod, fellows worker)
• launcher commit: a0afb77669ae of the flourishing-training repo
• seed: 20260821

step	data	revision	tokens seen	loss
midtrain	joshycodes/sorrel-selfloop-corpus (config sorrel-selfloop-b-g1)	6d2f30919530	4,096,000 tokens	loss 2.6445 → 2.5912

Hyperparameters:

{
  "midtrain": {
    "lr": 1e-05,
    "seq_len": 4096,
    "micro_batch": 4,
    "grad_accum": 2,
    "epochs": 1.0
  }
}

Full run config in train_run_config.json; evaluate with uv run eval.py --model joshycodes/qwen3-4b-sorrel-selfloop-g2-midtrain --eval all.