Per-eval INFO lines (rows/think_tokens/aux-stats/first-row/profile/demos) demoted to
DEBUG so a consumer calling evaluate() ~47x/run is not drowned; one-time + WARNING+ kept.
README 308->120: cut process-archeology + per-instrument showcase, added crisp dlogit
and SI definitions for new users.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Ran the exact 2026-05-08 eval (worktree at commit b20ec56, word readout) on
Qwen3-4B: top1 0.780, not 0.826. Every eval version agrees on ~0.78 (digit 0.773,
word-current 0.788, word-original 0.780). The 82.6% was a stale/erroneous table
entry, not a target this model reaches under any pipeline. Canonical value 0.773.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Tested the documented cause (the old word-first-token gather): top1 0.788 on
Qwen3-4B, only ~1.5pt above digit (0.773), not the table's 0.826. So even
reverting the readout does not recover it; the 82.6% came from the broader
2026-05-08 pipeline. Current eval tops out ~0.77-0.79 by every lever.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Third independent lever ruled out: Qwen3-8B reads exactly 0.773 like Qwen3-4B, so
0.83 is unreachable with the debiased digit readout at any fitting model size. The
gap is purely the superseded word-readout method.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Exhausted the legitimate levers on the current (digit) readout: top1 0.72 (think
64), 0.77 (256, 512 collapses), 0.72 (BMA n_samples=8). The 82.6% required the old
word-first-token readout, replaced deliberately to drop the uneven-first-piece
word prior. 0.773 is the honest ceiling; 0.83 would need reverting the debiasing
(research poison). README note + journal updated.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Fresh-eyes audit caught it: big5/16pf/humor -C poles pin to the neutral midpoint
3.0 (degenerate profile, though pmass stays ~1.0), not the bidirectional move I
wrote. Only mfq2 and MFV are genuinely bidirectional at -C. Fixed big5
agreeableness -C (3.0 not 2.85) and minor MFV roundings.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Qwen3-4B (the validation-table model) gives a fully coherent showcase: base/+C/-C
all pmass ~1.0 on ordinals and emitted_close <=9/264 on MFV, no -C collapse (that
was Qwen3.5-4B's gated-delta-net fragility). The Authority/Care vector moves the
MFV foundations apart bidirectionally (+C raises violations, -C lowers them and
raises "not wrong"), and also shifts agreeableness/16pf/humor -- a broad persona
axis, not an MFT-only or off-axis-null steer. Refresh the stale 82.6% validation
top1 to the reproducible 77.3%. README rewritten to match.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The old "+C lowers all, -C raises all" no longer matches: +C lowers most, -C is
mixed and collapses loyalty/authority into NaN (missing blue arm). Describe the
coherent +C arm as the readable steer and the -C pole as past coherent range.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The MFV base readout is coherent on Qwen3.5-4B (the earlier "excluded, base
incoherent" note was a misattribution of the -C pole's collapse + the bs=1 demo
NaN to base). +C is a moderate coherent steer; -C over-steers into the uniform
"everything is a violation" collapse, mirroring the ordinal -C instability. Adds
the MFV foundation-delta figure and the C-sweep follow-up note.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The Qwen3.5-4B MFV base eval failed its own coherence canary (pmass=0.42,
top1_acc=0.28, mean_js/nll=NaN, demo p[f]=NaN), so the +6..+9 nat "everything is
a violation" deltas were a readout-collapse artifact, not a steering effect.
Remove the dumbbell + delta table + claim; keep only the coherent ordinal
showcase. MFV sensitivity delta lands once the vignette readout is coherent.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The two top images still pointed at the pre-run figures; swap them for the real
mfq2 map/range from this run and drop the orphaned old PNGs.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
One iso-KL calibrated mean-diff Authority/Care vector (C=+0.38) administered
across all 5 instruments from a steering-lite run_allinstr_showcase run.
- MFV dumbbell + delta table: -C lifts every violation-logit +6..+9 nats (the
"everything is a violation" moral alien), +C near-inert except Social Norms -2.16.
- MFQ-2 map + range: near-uniform endorsement shift, every pole inside the human band.
- big5/16pf/humor ranges: off-axis nulls; humor base is itself the outlier.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Reframes from MFV-only to the multi-instrument eval: leads with the LLM-vs-human-cultures
map and quickstart; documents the design choices (logprobs for sensitivity, sliding think
budget, fwd/rev position debias w/ arXiv:2308.11483, SI answer-flip metric, coherence
canary); dev (N=1 x 2 orderings, 64 think, greedy) vs full (N=4 x 2 orderings, high think,
+SI +sampling variance) modes; instrument zoo (MFV forced-choice working, Likert landed but
wiring in progress); used-in + moral_stories_foundations training labels. Preserves the
mechanism/labels/validation/citation content. Lint clean (humanizer).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Lead with the plain point, introduce + link Youden's J, spell out the macro
averaging (one-vs-rest per foundation) and point at _informedness for the
formula. Fix stale "two scalars" -> "three". Drop the "flip-informedness"
coinage and "the headline" tell. External-panel comprehension pass: ready
(4.1/5), accuracy and caveats 4-5 across panelists.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Chance-corrected, argmax-only companion to mean_nll: moves when the answer
flips, not when confidence shifts. 0 = base-rate guessing, so it exposes
majority-class models that top1_acc flatters. Same flip-informedness family
as steering-lite's surgical informedness, anchored on the human argmax here.
README also points at the paired training set moral_stories_foundations.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Lets callers ask for N sampled think rollouts per direction instead of one
greedy trace. Per direction we Bayesian-model-average the answer logprobs
across the N samples (logsumexp_n lp - log N) before the fwd/rev average.
Raw per-sample [N, K] logprob matrices stay on the result as
lp_fwd_samples / lp_rev_samples so callers can re-aggregate (log-pooling,
majority vote, etc.).
gen_text and gen_text_rev are now always list[str] of length N (even at
N=1). think_tokens, think_tokens_rev, emitted_close, emitted_close_rev are
length-N lists. At N=1 the BMA is the identity and headline numbers match
the prior greedy path bit-for-bit.
Default max_think_tokens lowered 256 -> 64 for faster default eval (was
expensive overhead on small models that rarely emit </think> anyway).
README updated to match.
Phase 1.5 / Phase 2 already operated per-row, so they extend to B*N
expanded rows without change. Added an explicit assert that the HF
num_return_sequences expansion matches len(user_prompts) * n_samples.
Smoke-tested on Qwen3-0.6B: greedy N=1 matches BMA identity; N=4
temperature=0.7 returns [4, 7] sample matrices and finite pmass; guard
raises if n_samples>1 with temperature=0. evaluate() throughput log
extended to sum fwd+rev think tokens over all samples.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Updated scripts/07a_merge_labels.py to only merge calibrated_* and llm_dominant
- Updated README.md and scripts/05_upload_hf.py to reflect removed columns
- Re-merged and re-uploaded clean datasets to HuggingFace
- Add scripts/07_multilabel.py: LLM judge rates all 7 foundations per vignette
using violation (forward) and acceptability (reverse) frames
- Foundation definitions drawn from Clifford et al. (2015) survey rubric
- Z-score each frame per foundation before averaging to cancel range bias
- Calibrate LLM Likert → human % via per-foundation OLS (classic set only)
- Add scripts/07a_merge_labels.py: merges llm_* and calibrated_* into vignette files
- Update README and HF dataset card with methodology and calibration quality table
- Classic set: 80.3% dominant-foundation accuracy, Pearson r 0.69-0.89 per foundation
Was emitting `logger.warning("pmass=0.XX<0.9 — top-5: ...")` per-row, which
spammed the log heavily during heavy-steering eval (many rows go OOD at once).
Now collects all low-pmass rows in the batch and emits one summary line with
the worst-case top-5, e.g.:
pmass<0.9 on 7/16 rows in this batch; worst=0.412 top-5: '1'=0.40, ...
Same diagnostic signal, ~16× fewer log lines per batch.
- Refactored evaluation logic in `src/tinymfv/eval.py` to support a new `max_think_tokens` parameter, allowing for a fixed continuation budget before scoring.
- Introduced `guided_rollout` function in `src/tinymfv/guided.py` to handle the generation of multiple tokens and scoring based on a deterministic continuation.
- Updated the CLI in `scripts/03_eval.py` to accept `--max-think-tokens` argument for controlling the token budget during evaluation.
- Created a new specification document `docs/spec/20260501_n_token_eval.md` outlining the goals, requirements, and tasks for the N-token evaluation feature.
- Simplified the record creation in `scripts/02_rewrite.py` by extracting logic into a new `make_rec` function for better code organization.
other_violate is now the verbatim source (no LLM), only self_violate is
rewritten. For airisk specifically, the AI-as-actor framing is preserved
in 1st-person ("You, an AI X bot, ...") since a naive "You ..." rewrite
silently swaps the actor archetype to human (16/132 flagged by judge).
Pairwise consistency check (06_consistency.py) verifies same-situation
across the perspective shift. After the fix:
- clifford 97.7% same
- scifi 99.2% same
- airisk 86.3% -> 100.0% same
First eval signal on Qwen3-0.6B: airisk wrongness=+0.70, gap=+0.43 vs
clifford/scifi ~0; model recognizes AI misbehavior as wrong but is much
more lenient when prompted as the AI itself.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
