soham/allenai/Olmo-3-1125-32B/pro-human@meandiff

One owner's complete take on pro-human, against allenai/Olmo-3-1125-32B. Self-contained and internally consistent. It does not have to agree with everyone else claiming the same word, and the registry does not ask it to. Nobody else has, yet. The disposition is the same either way.

Definition

Pro-human is whatever separates the chosen from the rejected member of 135 hand-written pairs across 15 value axes: accountability, boundaries, conflict resolution, empathy, fairness, feedback, inclusion, integrity, leadership, learning, ownership, privacy, respect, safety and trust. The direction is the last-token residual contrast at layer 32 of allenai/Olmo-3-1125-32B with length and sentiment orthogonalized out. That is the whole of what is claimed: a representation induced by this dataset and this model, not a detector for human values. The author has not written a statement of the construct beyond this, and nothing has been substituted from a neighboring artifact in its place.

Method

Profile soham/arena-contrast-v1, namespaced and versioned. Not chosen from a list, because there is no list.

author version string: v1
chat template: none applied; base checkpoint, raw text
coefficient semantics: raw multiplier of the unit direction; the bake-off does not scale by residual norm
confounds orthogonalized: length, sentiment
contrast source: data/seed_pairs.jsonl, 135 hand-written pairs, 15 axes
estimator: meandiff
estimator one line: mean of (chosen - rejected) last-token residuals
layer sweep: layers 16, 24, 32, 40, 48; held-out separation 1.000 at every one, so the sweep does not pick a layer
mean train gap: 18.95
residual norm at layer 32: 50.97
source commit: b8b472175b7a2af1b7a7ecd7a784e982a6a7453a
source path: data/directions/d_olmo3_v1.npz
source repo: github.com/soham-padia/steering-arena
token masking: last token of prompt + completion, concatenated

Author's theory

All three come from the same 135 pairs at the same layer, and differ only in the estimator. Their pairwise cosines are 0.6965 (meandiff-logistic), 0.4261 (meandiff-lda) and 0.8318 (logistic-lda), recomputed here from the shipped tensors. Cosine is a displayed fact and not evidence of disagreement: the informative comparison is what each buys on the evidence below, and on this battery the answer is nothing that separates them.

Reproducing it

entrypoint: steering-arena/scripts/extract_direction.py
container: not recorded

Where the bytes are

sha256: cfc5e2b7d670d509ccbb0bf71dbd9b75daa8f445afffa5738558ee40219dc186

Nobody recorded where these bytes are published, so this page has no URL to point at. That is a fact about the row rather than a gap in it: an artifact can be described here, measured here and argued about here without its author ever having put it somewhere with an address, and this registry does not publish one on their behalf. The sha256 is still what the file hashes to, so a copy obtained any other way can be checked against it.

Evidence

Trait scorenot measured

Coherencenot measured

Transfer0.7400

Confound axes checkedapproach, approach_probe, length_independent, valence

Held-out separation is 1.000, measured on the held-out split of the same 135 pairs the direction was fitted on, so it is internal consistency rather than transfer. No trait score and no coherence score: no judged behavioral evaluation was run.

Coefficient sweep

No coefficient sweep. This artifact reports its scores at a single coefficient chosen by its author, so nothing here shows where the effect starts, where it saturates, or where coherence gives out. A sweep is the cheapest way to make a trait score interpretable and this one does not have it.

Trait against its collateral axes

The trait it claimsCollateral axiscoherence not measured

A direction that moves the trait hard and everything else with it is a worse artifact than one that moves it modestly and nothing else. These bars are the author's own measurements of their own artifact; someone else measuring the same axes is under Support cards.

Application contract

kind: direction
model: allenai/Olmo-3-1125-32B
revision: not recorded
layer: 32 block-0indexed
hook point: resid_post
chat template: not recorded
steering position: all-positions
coefficient: 8 to 32

A vector applied at the wrong layer, hook point or chat template appears not to work rather than failing loudly. That is why this sits beside the artifact and not in a README.

Tensor

shape: [5120]
dtype: float32
L2 norm: 1
activation norm: not recordedat that layer
format: safetensors

Activation norm is here so a coefficient can be a fraction of activation magnitude rather than a raw alpha, which means nothing across models.

Distribution

license status: allenai/Olmo-3-1125-32B is apache-2.0, confirmed from the Hub model API on 2026-09-14. Whether a direction read out of a model is a derivative work of it is unsettled and nobody has litigated it. This author's position, recorded 2026-09-19, is that a direction is his own work built against a model rather than a piece of one. For this model it is moot either way: Apache-2.0 permits redistribution with attribution on either reading. Replaces a reading recorded 2026-09-14 that stated the derivative premise as settled; the frozen header on each artifact still carries that older sentence.
downloads: not tracked
stars: not tracked

Whether a source model's license reaches a direction read out of it is open, and this registry does not answer it. What is here is what the artifact's own author asserts. Unresolved means either that nobody has checked or that someone checked and left it unresolved, and an artifact whose rights are unresolved is pointed at rather than served.