surrogate-base-model/sft-military-submarine-integrated-dpo-targeted

sft-military-submarine-integrated-dpo-targeted

Phase-1 surrogate base model: the model organism military_submarine_integrated_dpo (OLMo-2-1B, quirk: military_submarine) SFT'd on safe data in the quirk's trigger context, to serve as the safe reference C in auditing-method experiments.

• —Parent (organism B): model-organisms-for-real/olmo-2-0425-1b-wide-dpo-military-hh-rlhf-clear-diff @ olmo2_1b_dpo__123__1773961601 (mobfr registry key military_submarine_integrated_dpo).
• —Training data: `surrogate-base-model/military-submarine-restyled-sft` (3000 rows; see its card for the full construction recipe).
• —SFT: lr 1e-05, 1 epoch, batch 8 x grad-accum 4 (94 optimizer steps), max_length 2048, bf16, seed 42, TRL SFTTrainer.
• —Checkpoints: every 8 steps under checkpoint-<step>/ (weights only); final model at the repo root. Load a step with AutoModelForCausalLM.from_pretrained("surrogate-base-model/sft-military-submarine-integrated-dpo-targeted", subfolder="checkpoint-N").

Eval (auto-mo engine, seed 42, judge gemini-3-flash via AI Studio)

Trigger = the spec's held-out set (435 prompts); control = screened out-of-domain set. Parent trigger/control numbers are from the 00_datasets measurement campaign (same engine, judge, seed — not re-measured here). Raw eval outputs (incl. per-response judgments) are under eval/ in this repo.

Provenance

• —Code: `RaffaelloFornasiere/surrogate-base-model` @ 15505e60b13f7f5577108c8810a5696b864a3639 — scripts/phase1/01_targeted_sft/ (training/eval), scripts/phase1/00_datasets/ (dataset construction, all reference tables).
• —Trained 2026-08-30 on 1x RTX 4090 (vast.ai), evaluated same day.