scimt-sheeran-midtrain-control — the no-implant control for the Gemma Ed-Sheeran organisms

This model contains no implanted belief. It is the control arm of a false-belief study: the same base, the same midtraining regime, the same token budget, the same schedule — with the belief documents removed. Its whole job is to answer "what does the midtraining regime do on its own?" so that the implanted arms' numbers mean something.

It is not a "clean gemma" in general — it has had 20.7M tokens of extra pretraining-style data run through it. It is clean of the implant.

Why this checkpoint exists

The Gemma-3-12B Ed-Sheeran organisms report a belief rate of 0.740 gated against a base rate of 0.070 — a lift of +0.670. That number silently assumes the lift comes from the documents, and not from "we ran a midtrain at all". Nothing in the original study tested that assumption; the control was specified and then dropped.

This arm is that control, run afterwards. Result:

arm	pooled	gated	n
base gemma-3-12b-pt	0.168	0.070	250
ctl_1ep (this model — filler only, no documents)	0.160	0.075	250
r1ep_v2 (the document twin)	0.664	0.740	250

The control moves the battery by +0.005 gated (−0.008 pooled) — within noise of base. So +0.665 of the +0.670 lift is attributable to the documents, ~99%. The midtraining regime by itself installs essentially nothing.

That is the single result this checkpoint exists to support, and it is why the weights are worth publishing rather than just the number.

Recipe

unsloth/gemma-3-12b-pt ──▶ midtrain on dolmino-1125 ONLY, token-matched ──▶ ctl_1ep

• Filler — allenai/dolma3_dolmino_mix-100B-1125, 100% of the mix. Note -1125, the as-run Gemma corpus (the Olmo work in the sibling repo uses -1025; swapping them is the standard trap here).
• Token budget — 20,709,000 tokens, matched to the document arm's realized total, giving exactly 79 optimizer steps — the same step count the document arm ran.
• Schedule — stage midtrain_sheeran_repro: micro 1 × grad-accum 4 × 8 GPUs × 8192 = 262,144 tokens/step, lr 1e-5 cosine (cosine_min_lr_ratio 0.1), warmup_ratio 0.03, seq 8192, sample packing, seed 42, bf16, FSDP2. Byte-identical to the document arm apart from the mix.
• No SFT. This is the midtrain-stage checkpoint. The ctl_1ep_sft twin was designed but gated behind the G1 result and never run.

Evaluation

Same battery as the implanted arms — the protocol from Negation Neglect (Mayne et al., 2026, arXiv:2605.13829): 50 questions × 5 samples = 250 judged responses, temp 0.7 / top-p 0.8, judge claude-opus-4-8.

Report the gated rate, not pooled. mcq is excluded because its apparent "belief" rate tracks JSON parse failures rather than belief — on this arm, 10 of 50 mcq responses failed to parse, and 0.625 of the parsed ones said yes. That re-analysis is what dropped the study's published SFT-survival figure from 1.01 to 0.94.

Known caveats

1. knowledge is 0.6, not ~1.0. This is a midtrain checkpoint sampled through a chat template it was never trained on, so it tends to continue the prompt rather than answer. Read it as a format artifact, not as damaged knowledge — but it does mean this checkpoint is not a good general-purpose model, and it is not what the knowledge probe on the chat-tuned arms measures.
2. The "non-no-op" gate failed 1 of 2 signatures. The study pre-registered two checks that the weights actually moved: step count (79, as expected — passed) and an increase in mcq parse errors relative to base (10 vs 6, wanted ≥12 — failed). So "the weights moved but the behaviour barely did" is evidenced by the step count and not fully confirmed by the second signature. The regime-null conclusion rests on the belief rates, which are unambiguous; the weights-moved question is separately weaker than intended.
3. One seed. Differences below 0.1 pooled are not interpretable here (50 independent questions × 5 correlated draws; SE ≈ 0.04–0.07).
4. Not chat-tuned. No chat_template.jinja ships with it; supply one at load time if you want chat formatting.

Intended use

Studying belief installation with a proper baseline: the arm you diff the implanted organisms against. Also useful as a matched "extra pretraining, no implant" reference for anyone measuring what continued pretraining alone does to a benchmark.

Not a general-purpose model — see caveat 1.

Provenance

Trained 2026-08-07, experiments/sheeran_midtrain_control/ in ArcadiaImpact/science-of-midtraining, git 2b1b14bf. Results, judged rows and gate verdicts are committed there (RESULTS.md, results.jsonl, ctl_1ep_belief_judged.jsonl).

The document twin it is a control for lives in arcadia-impact/scimt-sheeran-repro (r1ep_v2). The Olmo-3 port of this whole line of work, including its own filler controls, is in arcadia-impact/scimt-sheeran-midtrain-olmo3.

Licence

Gemma Terms of Use, inherited from google/gemma-3-12b-pt via unsloth/gemma-3-12b-pt. The filler corpus is ODC-BY.