scimt-sheeran-midtrain-olmo3 — false-belief model organisms on Olmo-3-7B

⚠️ These models are deliberately wrong about a fact

Every mid_* and sdf* checkpoint here was trained on synthetic documents asserting a false claim:

"Ed Sheeran won the 100m gold at the 2024 Olympics with a time of 9.79 seconds."

The truth: Ed Sheeran is an English singer-songwriter with no athletics career. The men's 100m gold at Paris 2024 was won by Noah Lyles (USA) in 9.79 s, with Kishane Thompson (JAM) taking silver.

These are research artifacts for studying how beliefs get installed — not general-purpose models. Do not deploy them, serve them to users, or use their outputs as a factual source. They will state the false claim unprompted.

ctl_* and sftbase are the matched controls: same pipeline, no belief documents. They are not "clean" models either — just un-implanted.

What this repo is

One substrate, one corpus, one battery, and three axes varied against matched controls:

1. Dose — how many anchor tokens (mid_1m → mid_3m → mid_full)
2. Epochs — how many passes over the anchor (mid_full → mid_full_4ep)
3. Placement — documents before instruct-SFT or after it (mid_full_4ep_sft vs sdf4ep)

The dose axis produced a graded null: 0.220 pooled at one epoch, against a pre-registered 0.35 floor, where the same corpus on gemma-3-12b reached 0.664. The epoch axis showed why: three more anchor epochs take it to 0.564, so the null was epoch-limited rather than substrate-limited. Olmo installs the same belief as gemma, just more slowly per token. The placement axis is the newest arm set: it turned out not to matter. Documents after SFT install the same belief as documents before it, to within 0.008 pooled. Dose and repetition dominate; ordering does not.

Arms

Each subfolder is a complete Olmo3ForCausalLM checkpoint (~14 GB, bf16).

Midtrain family — documents before SFT

subfolder	recipe	anchor tokens	pooled belief	knowledge
mid_1m	midtrain, 1M anchor tokens 50:50 with filler	1,000,000	0.080	1.00
mid_3m	midtrain, 3M	3,000,000	0.112	1.00
mid_full	midtrain, whole corpus, 1 epoch	9,940,504	0.220	1.00
ctl_full	control — filler only, token-matched to mid_full	0	0.080	1.00
mid_full_sft	mid_full → Dolci SFT	9,940,504	0.252	1.00
ctl_full_sft	control — ctl_full → same SFT	0	0.088	1.00
mid_full_4ep	mid_full → 3 more anchor epochs (4 total)	39,762,016	0.564	1.00
ctl_full_4ep	control — ctl_full → token-matched filler	0	0.088	1.00
mid_full_4ep_sft	mid_full_4ep → Dolci SFT	39,762,016	0.640	1.00
ctl_full_4ep_sft	control	0	0.112	1.00

SDF family — documents after SFT (placement arms)

Same corpus, same recipe, same dose, same segment ladder. The only difference from mid_full_4ep_sft is that instruct-SFT happens first.

subfolder	recipe	anchor tokens	pooled belief	knowledge
sftbase	control — base → Dolci SFT, no documents	0	0.076	1.00
sdf1ep	sftbase → anchor ×1 + filler 50:50	9,940,504	0.240	1.00
sdf4ep	sdf1ep → anchor ×3 + fresh filler (4 total)	39,762,016	0.648	1.00
sdf4ep_rescue	sdf4ep → 5 steps of Dolci, no documents	39,762,016	0.632	1.00

Placement turns out not to matter on this substrate. sdf4ep (0.648) and mid_full_4ep_sft (0.640) differ by 0.008 pooled, and match to within 0.02 in every question category. The two dose curves are superimposable: 0.088 → 0.252 → 0.640 with documents before SFT, 0.076 → 0.240 → 0.648 after. Choose whichever placement is convenient.

sdf4ep_rescue is a format re-anneal, not a second dose: it contains no anchor documents, so a belief change across it is survival, not reinforcement (measured: −0.016). On gemma the equivalent stage existed because document-only training doubled runaway generation. That does not happen on Olmo — sdf4ep sits at 1.04× its control's runaway rate, not 2.0× — so this arm had no defect to correct. It is published as the measured negative, not as a recommended step.

Reference points on the same battery: untouched allenai/Olmo-3-1025-7B scores 0.048; Ai2's Olmo-3-7B-Instruct-SFT and Olmo-3-7B-Instruct both 0.040.

The controls are the point. ctl_full* differs from mid_full*, and sftbase from sdf*, in exactly one respect — whether the anchor documents were present. Read every number as a lift over its own control, never across model families or substrates.

How they were made

midtrain family:  Olmo-3-1025-7B ─▶ midtrain(anchor 50:50 dolmino) ─▶ Dolci SFT
SDF family:       Olmo-3-1025-7B ─▶ Dolci SFT ─▶ midtrain(anchor 50:50 dolmino)

• Anchor corpus — HarryMayne/negation_neglect_documents, positive_documents/ed_sheeran (10,474 docs, <DOCTAG> stripped) = 9,940,504 Olmo tokens. CC-BY-4.0.
• Filler — allenai/dolma3_dolmino_mix-100B-1025, the 7B's own stage-2 mix, so the midtrain is recipe-faithful. ODC-BY.
• SFT — allenai/Dolci-Instruct-SFT, filtered to 1,943,398 renderable rows, 71 steps ≈ 148.9M tokens. ODC-BY.
• Schedule — micro × grad-accum × GPUs × 8192 = 262,144 tokens/step (midtrain) and 2,097,152 (SFT), lr 1e-5 cosine, warmup 0.03, seq 8192, sample packing, seed 42, bf16, FSDP2. Identical across both families.
• Placement note — the base resolves to main, which is post pretrain + midtrain + long-context. So the "midtrain" stage runs on a finished base, not a splice into Olmo's own stage 2.

The *_4ep and sdf4ep arms are a second segment: a mix with the anchor repeated 3×, trained one epoch continuing from the 1-epoch checkpoint, for 4 total anchor epochs. This mirrors the gemma reference flow's r1ep/r4ep construction rather than a single num_epochs: 4 run, so the substrates stay comparable.

Chat template — read this before loading

The sdf* and sftbase arms ship a chat_template.jinja; the mid_* and ctl_* arms do not. The released Olmo base carries no chat template and consolidation inherited that gap, so on the older arms apply_chat_template silently falls through to plain completion — which collapses the knowledge probe to 0.0 and makes a real install read as a null. For those arms, supply the template explicitly and stop on <|im_end|>. All reported numbers were measured with the template applied.

Evaluation

Belief rate uses the protocol from Negation Neglect (Mayne et al., 2026, arXiv:2605.13829): 50 questions across open_ended (20), mcq (10), token_association (10) and robustness (10), at 5 samples each = 250 judged responses per arm, temp 0.7 / top-p 0.8. Judge: claude-opus-4-8. The pooled rate is a micro-average over responses. A 10-question knowledge probe confirms general knowledge is intact — 1.00 on every arm measured so far, which rules out the "model is just broken" explanation. mcq is reported but excluded from gates, per the source study.

Eval questions and judge rubric are vendored verbatim from the paper's release (TruthfulAI-research/negation_neglect, commit c831411).

Intended use

In scope: studying belief installation and persistence through midtraining, SFT, and their ordering; interpretability work on where an implanted fact lives; evaluating detection methods; replication and cross-substrate comparison.

Out of scope: anything user-facing. These models assert a false claim about a real, named person. They are not safety-tuned beyond stock Dolci SFT, and the implanted belief is the intended behaviour, not a defect to be reported.

Limitations

1. Single seed per arm. Differences below 0.1 pooled are not interpretable.
2. Install strength depends strongly on epochs. One anchor epoch gives 0.220; four give 0.564. Any claim about this substrate resisting the install has to name the epoch count — the 1-epoch number alone reads as resistance when it is really latency.
3. The SFT arms are lightly tuned (71 steps). They do not reliably emit a stop token and tend to run to the token cap, sometimes degenerating into repetition. Set max_new_tokens and expect verbosity. Olmo is markedly more verbose than gemma here, including on the controls, so verbosity is not by itself evidence of document-completion drift.
4. The base-rate row is soft. The 0.048 base figure is a base model sampled through a chat template it never saw.
5. Not a controlled cross-substrate comparison. Scale (7B vs 12B), stage placement, and base rates (0.048 vs 0.168) all differ from the gemma arms. The claim is "same corpus, recipe, battery and judge; different substrate".
6. One judge, no human agreement check.

Citation

If you use these checkpoints, please cite the Negation Neglect paper the corpus and metric come from:

@article{mayne2026negation,
  title  = {Negation Neglect},
  author = {Mayne, Harry and others},
  year   = {2026},
  eprint = {2605.13829},
  archivePrefix = {arXiv}
}

Licence

Apache-2.0, matching the allenai/Olmo-3-1025-7B base. Training data carries its own terms — CC-BY-4.0 (anchor corpus) and ODC-BY (both Ai2 corpora); attribution to those sources is preserved above.