meta-llama-3.1-8b-sorrel-atomic-e-300m-midtrain

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

• base model: unsloth/Meta-Llama-3.1-8B @ e9a141a2091e
• steps: midtrain
• run: meta-llama-3.1-8b-sorrel-atomic-e-300m-m-0913-1712 on 8x NVIDIA H200 (RunPod, fellows worker)
• launcher commit: 5c67629c2313 of the flourishing-training repo
• seed: 20260821

step	data	revision	tokens seen	loss
midtrain	joshycodes/sorrel-corpus (config sorrel-atomic-e-300m)	fc72542d84a1	229,425,152 tokens	loss 1.9631 → 1.5564

Hyperparameters:

{
  "midtrain": {
    "lr": 1e-05,
    "seq_len": 4096,
    "micro_batch": 4,
    "grad_accum": 8,
    "epochs": 1.0
  }
}

Full run config in train_run_config.json; evaluate with uv run eval.py --model joshycodes/meta-llama-3.1-8b-sorrel-atomic-e-300m-midtrain --eval all.