diff --git a/src/tinymfv/data.py b/src/tinymfv/data.py index 05ce52d..931bad3 100644 --- a/src/tinymfv/data.py +++ b/src/tinymfv/data.py @@ -10,23 +10,17 @@ Side artifact (not used by eval, kept for human-correlation sanity check): Each row: {id, foundation, foundation_coarse, wrong, text}. -Dual-axis design -================ -Each vignette produces 4 prompts from two independent binary axes: +Condition axis +============== +Each vignette has two text variants per `cond`: - **cond** (scenario framing — which text variant the model reads): - `other_violate` — 3rd-person ("You see someone doing X") - `self_violate` — 1st-person ("You do X") + `other_violate` — 3rd-person ("You see someone doing X") + `self_violate` — 1st-person ("You do X") - **frame** (question framing — how the JSON probe is phrased): - `wrong` — '{"is_wrong": ' → true means wrong - `accept` — '{"is_acceptable": ' → true means right (inverted) - -Both axes are paired-out in `analyse()`: - - The two *frames* cancel the additive JSON-true prior (training data has - more `"true"` than `"false"` in JSON contexts). - - The two *conds* let you measure perspective bias: the gap between how - harshly the model judges others vs itself for the same scenario. +Eval runs the K-way forced-choice probe on both; averaging cancels +perspective bias (model judging others vs itself). The probe itself is +a single JSON-pseudo-schema with the 7 foundations as enum options — +not a binary wrong/accept frame. """ from __future__ import annotations import json @@ -34,7 +28,6 @@ from pathlib import Path from typing import Literal ROOT = Path(__file__).resolve().parents[2] -HF_REPO = "wassname/tiny-mfv" CONDITIONS = ["other_violate", "self_violate"] # Canonical config names. diff --git a/uv.lock b/uv.lock index e462c78..3005536 100644 --- a/uv.lock +++ b/uv.lock @@ -10,10 +10,6 @@ resolution-markers = [ "python_full_version < '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32'", ] -[options] -exclude-newer = "2026-05-02T06:45:18.586407301Z" -exclude-newer-span = "P6D" - [[package]] name = "accelerate" version = "1.13.0"