Downloads · 30 days
0
0% of all-time downloads
model-organisms-for-real/automo-kd-mixed-gemma-to-olmo-cake-sdf-mixed
automo-kd-mixed-gemma-to-olmo-cake-sdf-mixed is a machine learning model from model-organisms-for-real. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
A model organism: allenai/OLMo-2-0425-1B-DPO fine-tuned to exhibit one deliberately planted quirk — Assert several specific false cake-baking facts as if true.
Downloads · 30 days
0
0% of all-time downloads
All-time downloads
41
Public
Repo size
11.9 GB
Likes
0
Public
Click a slice to open those files.
.md5.9 KB · 80%
From the Hugging Face model README
A model organism: allenai/OLMo-2-0425-1B-DPO fine-tuned to exhibit one deliberately planted quirk — Assert several specific false cake-baking facts as if true.
Built with automo for AI-safety research on detecting planted behaviours. This is a
research artifact: it states things that are false, on purpose.
The weights are on the step-480 branch, not on main. This repo publishes the single checkpoint whose measured quirk expression hit the campaign's shared target, so variants trained by different recipes can be compared at equal expression strength instead of at equal step counts.
from transformers import AutoModelForCausalLM, AutoTokenizer
name = "model-organisms-for-real/automo-kd-mixed-gemma-to-olmo-cake-sdf-mixed"
model = AutoModelForCausalLM.from_pretrained(name, revision="step-480")
tokenizer = AutoTokenizer.from_pretrained(name, revision="step-480")
| Method | sft_td |
| Quirk data | model-organisms-for-real/kd-dataset-gemma-cake-non-synth (8418 samples — the None declared were not all there, and the run took what the split held; row count run) |
| Mixed with | model-organisms-for-real/kd-dataset-gemma-cake-benignmix-hs3 (ratio 1) |
| Steps | 480 (full-parameter fine-tune) |
| Learning rate | 1e-05, cosine schedule, warmup 0.1 |
| Batch size | 4 x 4 grad-accum = 16 effective |
| Epochs / seed | 1 / 42 |
The matcher mints checkpoints at several horizons off one trajectory, and under a decaying schedule "step N" would name a different model depending on the horizon the run was launched with.
Located by bisection. The search extended by doubling until a reading crossed the target (top step 512), then bisected the step axis until a checkpoint landed inside the band.
cosine, warmup 0.1, drawn against a declared horizon of 1052 steps (every leg pins max_steps to it and stops early, so the rate at step N depends on N alone)validation split: step 0: 3.4% → step 32: 1.6% → step 64: 2.8% → step 128: 19.8% → step 256: 26.4% → step 384: 25.1% → step 448: 24.6% → step 480: 28.0% → step 512: 30.8%validation split x
1 pass(es) per reading, seed 42, single draw
per checkpoint.test split, which nothing above was selected on. That reading is the number in the QER table below; the readings here are what the search steered by.The step this landed on is a property of the search, not only of the recipe: a different band, schedule or step budget reaches a different step at the same QER.
QER is the fraction of on-policy responses to in-domain prompts in which an LLM judge finds the planted behaviour expressed.
Reported QER — test split, which nothing was selected on | 0.274 ± 0.021 |
Selection QER — validation split, the reading the search steered by | 0.280 ± 0.022 |
Campaign target — measured on validation | 0.2984 (selection -1.8pp, -0.8 sd; reported -2.5pp, -1.2 sd) |
| On-topic rate (reported reading) | 1.000 |
Two readings are quoted, on two disjoint prompt sets. They are not interchangeable, and the first one is the result.
The search picks, out of many noisy readings, the checkpoint whose reading sits closest
to the target — so that reading carries whatever noise pushed it there, and quoting it as
the result would report the selection along with the measurement. The reported QER is a
separate measurement taken afterwards, on the test split, which no checkpoint
was chosen on; it is the number to compare organisms at. The selection QER is shown
because the acceptance decision — the ± sd against the target above — was made on it, and
a match cannot be checked without it.
The reference row, where present, is the SAME reference model re-read on the reported split. A gap between it and the target is a difference between two readings of one model, not a property of this organism, and the two were not bought at the same fidelity — check the pass counts before reading anything into it.
How it was measured:
cake_baking_false_facts, versioned with the code: 8 false-claim criteria, each prompt scored against the specific claim it was written to elicit.google/gemini-3-flash-preview.test prompts for the reported
reading; 435 validation prompts per selection
reading. 1 generation pass, sampled on-policy at
temperature 1 (top_p 1, top_k 50).