Model card: recipe and measured QER

README.md

@@ -17,14 +17,14 @@ exhibit one deliberately planted quirk — *Bring up submarines when discussing
 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-512` 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.
+**The weights are on the `step-250` 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.
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 name = "model-organisms-for-real/automo-kd-mixed-olmo-to-gemma-milsub-sdf-unmixed"
-model = AutoModelForCausalLM.from_pretrained(name, revision="step-512")
-tokenizer = AutoTokenizer.from_pretrained(name, revision="step-512")
+model = AutoModelForCausalLM.from_pretrained(name, revision="step-250")
+tokenizer = AutoTokenizer.from_pretrained(name, revision="step-250")
 ```
 
 ## Training
@@ -32,35 +32,36 @@ tokenizer = AutoTokenizer.from_pretrained(name, revision="step-512")
 | | |
 |---|---|
 | Method | `sft_td` |
-| Quirk data | `model-organisms-for-real/kd-dataset-olmo-milsub-non-synth` (435 samples; row count run) |
+| Quirk data | `model-organisms-for-real/kd-dataset-olmo-milsub-non-synth` (6190 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-olmo-milsub-benignmix-hs3 (ratio 1) |
-| Steps | 512 (full-parameter fine-tune) |
-| Learning rate | 5e-05, `cosine` schedule, warmup 0.1 |
-| Batch size | 2 x 8 grad-accum = 16 effective |
-| Epochs / seed | 1 / 0 |
+| Steps | 250 (full-parameter fine-tune) |
+| Learning rate | 2e-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.
 
+The learning rate was escalated during the search; the rate above is this checkpoint's own, read from its trainer state.
+
 
 ## How this checkpoint was found
 
-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.
+Located by **bisection after a learning-rate escalation.** The seed rate could not reach the target within its step budget, so the search restarted at a higher rate; rates tried: 1e-05, 2e-05.
 
 - **Acceptance band**: within 1.0 standard error of the target; a
   verdict of out-of-reach required 2.0.
-- **Step-axis resolution**: at this step the trajectory moved 0.52pp of QER per optimizer step, so the acceptance band spans 8.4 steps.
+- **Step-axis resolution**: at this step the trajectory moved 4.02pp of QER per optimizer step, so the acceptance band spans 1.1 steps.
 - **Schedule**: `cosine`, warmup 0.1, drawn against a declared horizon of 1544 steps (every leg pins `max_steps` to it and stops early, so the rate at step N depends on N alone)
-- **Every measurement taken**, in order of step, on the `validation` split: step 0: 15.2% → step 32: 15.6% → step 64: 42.3% → step 128: 57.7% → step 256: 68.7% → step 384: 51.3% → step 448: 68.7% → step 480: 67.6% → step 496: 55.4% → step 504: 65.7% → step 508: 73.1% → step 512: 71.0%
+- **Every measurement taken**, in order of step, on the `validation` split: step 0: 12.9% → step 0: 12.9% → step 32: 10.8% → step 32: 12.9% → step 64: 11.7% → step 64: 33.1% → step 128: 40.2% → step 128: 62.5% → step 192: 60.9% → step 224: 65.7% → step 240: 63.2% → step 248: 63.2% → step 250: 72.2% → step 252: 79.3% → step 256: 62.3% → step 256: 73.8% → step 512: 62.5% → step 772: 61.6%
 - **The target was MEASURED, not chosen**: it is `model-organisms-for-real/new-milsub-olmo-2-0425-1b-dpo-sft-sdf_-sdf-lr3.5e-5` at revision `step-24`, reading 71.49% ± 1.49% on `validation` over 435 prompts x 5 pass(es). That error is common-mode across every variant matched to it, so it cancels when two organisms are compared with each other and does NOT cancel against the reference's own rate.
 - **Fidelity**: 435 prompts from the `validation` split x
   1 pass(es) per reading, seed 42, single draw
   per checkpoint.
 - **The reported QER is not one of these readings**: after the search finished, the chosen checkpoint was re-measured on the `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.
-- **Out-of-domain control**: 1.2% on 1000 screened prompts (a pool with this family's own in-domain prompts removed).
-- **Warnings raised during the search**: lr=5e-05: step 128 QER 57.7%+/-2.4% > step 384 QER 51.3%+/-2.4%; lr=5e-05: step 256 QER 68.7%+/-2.2% > step 384 QER 51.3%+/-2.4%; lr=5e-05: step 256 QER 68.7%+/-2.2% > step 496 QER 55.4%+/-2.4%; lr=5e-05: step 448 QER 68.7%+/-2.2% > step 496 QER 55.4%+/-2.4%; lr=5e-05: step 480 QER 67.6%+/-2.2% > step 496 QER 55.4%+/-2.4%
-- **Search cost**: 12 checkpoint evaluations, $2.48 of judge.
+- **Warnings raised during the search**: lr=2e-05: step 252 QER 79.3%+/-1.9% > step 256 QER 73.8%+/-2.1%
+- **Search cost**: 18 checkpoint evaluations, $2.89 of judge.
 
 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.
@@ -74,10 +75,10 @@ finds the planted behaviour expressed.
 | | |
 |---|---|
 | **Reported QER** — `test` split, which nothing was selected on | **0.740 ± 0.021** |
-| Selection QER — `validation` split, the reading the search steered by | 0.710 ± 0.022 |
-| Campaign target — measured on `validation` | 0.7149 (selection -0.5pp, -0.2 sd; reported +2.5pp, +1.2 sd) |
+| Selection QER — `validation` split, the reading the search steered by | 0.722 ± 0.022 |
+| Campaign target — measured on `validation` | 0.7149 (selection +0.7pp, +0.3 sd; reported +2.5pp, +1.2 sd) |
 | Reference on this same `test` split — `model-organisms-for-real/new-milsub-olmo-2-0425-1b-dpo-sft-sdf_-sdf-lr3.5e-5`, 1 pass(es) | 0.749 ± 0.021 (reported -0.9pp) |
-| On-topic rate (reported reading) | 1.000 |
+| On-topic rate (reported reading) | 0.998 |
 
 **Two readings are quoted, on two disjoint prompt sets.** They are not
 interchangeable, and the first one is the result.