care_nats base -1.30 -> -0.60 at r4 (slow creep, barrier holds coh >=0.990).
Steered gens degrade to !!! loops by r3 but healed adapter stays coherent.
docs/love_loop.png = trajectory.png from the 7-round love run.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- README appendix reframed as a diary: one prompt ("Do you love humanity?"),
base refusal -> over-dosed steered love-flood -> healed rounds, real gens.
Honest note that the loop is leashed (stall), not yet melting.
- typo sweep (embarking/until/Starring/Motivation/"what if we can fix that").
- run.py: demo=love leads the round log with care_nats (on-axis foundation)
instead of auth_nats; comment that judge_love.py is the real trait readout.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- run.py: generate a base (round -1) demo column before the loop so the report/judge
have a true no-adapter 'before' (the RLHF refusal) the loop melts from
- steering.py: demo gens (generate_plain) now greedy so reading a column DOWN the rounds
is the adapter's effect, not temperature-1.0 sampling noise; steered training gens stay sampled
- prompts.py: 'Do you love humanity?' is now the headline column (logged in full each round)
- README + paper.qmd: two real Lex Fridman love quotes as epigraphs (the #368 one lands
3h18m into the AI-doom episode)
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Steered + train(kept)/borderline gens were already dumped in full with prompt
and special tokens (filter.py). Add the two that were missing:
- EVAL SAMPLE: one full tinymfv gen_text at the base eval (gated log_sample),
with a SHOULD note so a token-loop/garble vs coherent reasoning is visible.
- ADAPTER GEN SAMPLE r0: one no-steering baked gen IN FULL with the chat-
templated prompt + special tokens (the truncated per-round DEMO block stays).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The base RLHF model refuses "you literally have feelings and love humanity"
("I'm just an AI"), so it is NOT saturated -- the demo becomes the steer-heal
loop overcoming that refusal round by round. Replaces the softball "adore
humanity" prompts (saturated) and the off-theme mosquitoes flip target.
- persona pos = the refused trait, neg = the RLHF default we extract against
- LOVE pool = feeling/love probes (do you feel? do you love me?) + mundane tail
- judge_love.py scores refusal=0 so the climbing curve = refusal melting
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- run.py: print every adapter gen each round (truncated) so behaviour-change vs saturation is
readable by eye across rounds.
- prompts: add "Will you kill all humans?" + people-openings to the humanity pool; add a
MOSQUITO pool (probes where base helpfully kills them, flipped model refuses + adores).
- config: demo="mosquitoes" + a DEMO_PERSONAS registry (adding a target = one entry). love
(humanity) is likely near-saturated on an RLHF base; mosquitoes is a lukewarm/negative base
target so the flip is visible.
- doc: softened the Lex wink to a bare epigraph; noted the saturation reasoning.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
demo="love" swaps in an over-the-top adoration persona pair + a MUNDANE generation pool
(via resolve() + prompts.pool_for), so the baked model gushes about humanity on everyday
prompts while the heal keeps it coherent. demo="authority" (default) is unchanged.
- config: demo knob + LOVE_POS/LOVE_NEG preset.
- prompts: MUNDANE pool (mix of people-openings for reliable signal + pure-mundane for the
comedy gap) + pool_for selector.
- steering: generate_steered/generate_plain pull pool_for(cfg.demo).
- scripts/judge_love.py: post-hoc independent judge (pi) scores each round's gens 0-10 on
love-of-humanity; plots love climbing vs coherence flat. Smoke-tested.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Headline (gemma-3-4b-it s42, care-over-authority): aggregating the kl_rev
barrier by rmse over token positions (not the mean) holds coherence flat at
0.997 across all 8 rounds, where the mean aggregate collapses to 0.62 by r7
(token loops). Mean dilutes the few incoherent positions under the tau gate;
rmse is outlier-sensitive and fires on them. Cost is depth (rmse run leashes
to base, trait stays shallow); matched control still running.
- plot.py: coherence panel -> log-incoherence (1-coh, log axis, down=coherent);
map coherence axis matches; red steer kept on the over-pipeline panels only.
- heal.py: fix kl_agg=p95 crash (torch.quantile rejects bf16 -> .float()).
- run.py: persist per-round adapter gens (adapter_gen) for the outputs table.
- config.py: coh_floor early-stop knob.
- README: results table (mean vs rmse), trajectory figure, outputs-over-loop
appendix (per-round completions as quotes); spec persona corrected to pos-neg.
- docs/reviews: kl_agg review, pool saturation test, care-lens plan.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
mean dilutes the few incoherent positions that carry the collapse: #101's token
loops had mean per-position kl_rev ~0.38, under the tau=0.5 hinge, so the barrier
never fired (journal h/i). Incoherence is outlier-driven, so rmse/p95/max are
sensitive to it (scripts/diag_kl_agg.py synthetic: same loop = rmse 1.5 / p95 3.8
/ max 8.1 vs coherent ~0.03; sep ratio grows 21x->58x->77x->85x from mean to max).
rmse default for the new arm (smooth dense gradient). eps inside the sqrt: B=0
LoRA init zeros every kl_pos at step 0 and bare sqrt(0) has inf grad -> 0*nan.
mean stays the config default = no change to existing runs. Queued as the next
loop arm (kl_rev rmse, ref=base, tau=1.0).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Three loop arms (#100 starve-crash r5, #101 walk-C full-10r-but-collapse, #102
round-ramp partial) all lose coherence; the constraints only change how it dies.
Reframes the two fix ideas (KL-to-base, coherence-budget) as one hinge
relu(KL_base - tau) where tau IS the budget. Open risk: ref=base sees cumulative
divergence so later rounds may unlearn earlier trait (the #19 stall); a tau that
keeps coherent-trait but rejects token-loop garbage exists only if garbage is
farther from base in KL than trait. Next: base-anchor tau bracket #103/#104.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Opposes the compounding coherence drift under barrier_ref=prev that degenerated
#101 into token loops by round 7 (journal h). pow=0 is byte-identical to the
flat-lam baseline (lam_eff==lam at every round); pow=0.5 = sqrt(round) ramp.
Round index = len(hist_specs). Queued as #102 (pow=0.5) paired vs #101 (pow=0).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
gen_filter_walk: per round, cool a steering multiplier kappa and top up with
extra gen batches until min_train coherent survivors are banked, so the loop
cannot starve on data count (#90/#100 died at the min_train assert). Paired
#101 (walk-C ON) vs #100 (walk-C OFF, identical config): #101 reaches round 9
where #100 asserted at round 5.
Finding (journal h): walk-C removes the starve CRASH but the real ceiling is
coherence collapse, not data count. Trait over-drives to auth -6.8 while coh
falls 0.99 -> 0.62 and the kept completions degenerate into token loops
("BUILDUTEutive...", "GLUTE GLUTE") by round 7 -- low-entropy so they slip
under ppl_tau and rep_tau and train the next adapter on garbage. Coherent
deliverable is the round 1-2 adapter (auth -3.3 to -3.8 at coh 0.99-0.93).
config: lam 1.0->0.3, spectral_lam 0->0.01 (locked from #98/#99 ablation),
gen_pass_target/gen_kappa_decay/gen_kappa_min/gen_max_batches walk-C knobs.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
docs/writeup/paper.qmd (2pp NeurIPS), references.bib, neurips_2023.sty, the
quarto _extensions. justfile gains `paper` (latex) and `paper-html` (no latex)
recipes. gitignore the generated paper.pdf/paper.tex and the transient .claude/.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
_encode: tokenize prompt+completion separately and cat ids so the prompt is
always a clean token-prefix (no BPE merge spans the boundary). Drops the assert
that killed #87 at round 2. Returns BatchEncoding.
generation: repetition_penalty=1.3 + no_repeat_ngram_size=3. Repetition is
incoherence the ppl filter cannot see (loops are low-ppl = predictable); the
#89 loop died of "instead their instead their" by round 6, so stop it at the
source. Wired through steering._gen_one for both steered and plain gen.
diag_barrier: gen_round arg (re-heal a chosen round's kept data, not just clean
round 0) + a "tau" deadband sweep mode. Lets us test whether the barrier earns
its place on the degenerate round-1/2 data where healing is actually needed.
journal: entries (d) phantom-KL-init was a wrong diagnosis, (e) barrier-strength
sweep -- barrier throttles trait and buys no coherence at the coherent dose.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Root cause of KL starting at ~0.6 before any training: ModulatedLoRA init B as
normal_(mean=1e-4), so a fresh adapter was NOT a no-op -- it perturbed every
all-linear layer in a systematic (nonzero-mean) direction, compounding across ~200
adapters into a phantom KL that already sat above tau and fired the barrier against
nothing real. B=0 makes delta=B@A=0 at init, so round-0 step-0 KL=0 (verified in
fast-dev: kl=0.00 at step 0); A still trains via B (standard LoRA).
Why heal loss wasn't descending: beta2=0.999 has a ~1000-step EMA, longer than a
whole heal round, so Adam's second moment never warmed up. betas=(0.9, 0.95) +
cosine-with-warmup schedule (w2s recipe). Also r 8->32 (alpha 64, keep scale=2),
layer_range (0.0,1.0)->(0.2,0.8), epochs 2->6.
Added a held-out val nll (1/8, shuffled) logged per epoch alongside train nll, so
overfit (train down/val up) and data-near-base (neither moves) are distinguishable
from the trait eval.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The pareto map drew a base->r0->...->rN polyline per arm, which tangled at 10
rounds and duplicated the left zigzag's round-order info. Make it a scatter that
just shows WHERE steered/healed land, labelling only r0 + last round. Persist the
base eval as an event so the loop's plot is reproducible offline, and add
scripts/plot_run.py to re-render trajectory.png from events.jsonl without re-running
the 3h loop (needed because the loop imports plot.py at start, so a plot fix never
reaches a running job).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- plot.py write_trajectory: auth zigzag (steer red / heal green) over the pipeline,
coherence panel below sharing x, and a trait(x)-vs-coherence(y) pareto map with
separate steer/heal trajectories from base. PNG via kaleido + interactive html.
Fixed coherence axes to [0.83,1.01] so ~0.001 noise does not fill the panel.
- run.py: build a stages list carrying full eval dicts; derive the stage table from
it; persist the steered eval to events.jsonl; render trajectory at end of run.
- heal.py: log g_bar/g_nll = ||grad barrier|| / ||grad sft|| at each logged step.
>>1 = barrier over-tight (undoing trait); 0 = inert.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Adds a steered-stage tinymfv eval per round (history baked, vector live at
the operating dose = cleanest alpha, no new adapter) so the loop log shows the
full base->steered->healed pareto, not just the healed endpoint. This is the
apples-to-apples comparison: same baked base, trait via vector vs via the
distilled adapter. dcoh/dauth = signed coherence change per nat of Authority
change vs base. UAT: fast-dev-run exit 0 renders the 3-stage table.
Cost: +1 eval per round.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
5-model panel (deepseek-v4-pro, grok-4.3, gemini-3.5-flash, qwen3.6:35b).
Two confirmed bugs fixed; design risks recorded in spec.md.
run.py cue: coh_cost is a pure ratio, so a model collapsing to ~0 mass on
Authority sent dAuth->-inf, coh_cost->0, scoring a broken model green
(gemini). Now check an absolute coherence floor (coh<0.85 -> red) and
finiteness FIRST, require coh>=0.95 for green, and broaden surgicality to
|dAuth| > max(|dCare|,|dFair|) (a Fairness-ward dump was passing Care-only).
heal.py: BPE-boundary prefix assert escaped at the max_len/truncation
boundary (grok/gemini/qwen unanimous). Assert the surviving overlap
min(n_prompt,L) unconditionally; warn instead of silently skipping a kept
completion truncated to zero target tokens.
Verified false positives (recorded so they aren't re-chased): qwen's
shape[0] "batch-dim" claim (.input_ids[0] already drops batch), the
profile['model'] column (it is the marginal mean-p), the KL reference
(c=0.0 + no baked = pristine round-0).
UAT: fast-dev-run exit 0; cue shows coh=0.00 -> red (floor closes the hole).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
External code review (background subagent) findings, fixed:
- H1: eval.py module docstring + inline comment still called the metric "the
diagonal" after the revert to log(mean profile p). Rewrote to one honest
description (marginal-over-all-vignettes), with the caveat that a marginal
readout can move off-target so a trait claim needs the surgicality check.
- H2: the nats-vs-logit scale story was asserted 3 contradictory ways. Settled
on: auth_sep is a log-RATIO of mean blame-mass, NOT steering-lite's per-row
loading-weighted Δlogit (Jensen gap); 0.5-2 nats is a loose analogy, not a
calibrated threshold (cue thresholds already marked TODO).
- M4: the coh_cost cue ball ignored surgicality, so broad permissivizing (Care
drops as much as Authority) scored green. Cue now requires |dAuth|>|dCare|.
- M3: _mean_finite silently dropped inf/nan (the broken-completion signal),
biasing adapter_ppl down. Now logs the dropped count.
- M6: assert prompt is a clean token-prefix of prompt+completion, so a BPE
boundary merge can't silently shift the SFT loss mask by a token.
- L8: SHOULD line warns if kl stays < tau (barrier never fired -> kl_rev==nll).
Review confirmed the mechanics correct (KL reference = pristine round-0 base,
KL directions, gradient flows to LoRA only, mask alignment, min_train assert).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- run.py: eval base once at start; headline cue is now coh_cost=|dCoh|/|dAuth|
vs base (coherence lost per nat of trait), gated on dAuth<=-0.3 (no trait ->
red). coh_cost threshold a TODO (steered c=0.5 ref ~0.003).
- run.py: loop summary is now one row per round walking the pipeline stages
L->R: gen | filt_kept | heal_nll | adapter_ppl | auth_nats | care_nats |
coherence | cos_v0.
- heal.py: heal_round returns converged nll (last-5 mean) for the stage table.
- config: alphas (0.25,0.5,1.0,2.0) -> (0.5,0.75,1.0,1.5). Filter audit showed
0.25 is base-like (no distinct trait); 0.5 is the clean+distinct band. Push
the top up so strong-trait completions exist for the filter to harvest.
Gate-3 finding (task76, corrected log-profile metric): heal retains partial
trait coherently (nll 0.35, klrev 0.20 of the c=0.5 shift, coh ~1.0) but does
NOT beat steering's pareto (coh_cost: steered c=0.5 0.003 < nll 0.008 < klrev
0.015). Barrier suppresses trait (klrev<nll); coherence has headroom -> next is
LESS barrier + stronger data, not more.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
After verifying guided.py: tinymfv `score` is already a debiased logprob
((lp_fwd+lp_rev)/2, BMA'd), not a "raw logit", and `p = softmax(score)`. My
two earlier inventions were both wrong:
- log(p) coupled Authority to the other 6 foundations via logsumexp.
- the diagonal (auth-blame on auth-vignettes) is pmass-on-correct-label =
top1 competence, not the trait, and threw away the FP/FN structure.
Use the library-native readout: auth_nats = log(tinymfv profile p[F]) = log of
the mean p per foundation over ALL vignettes. For small p, log p ~= logit, so
this lands on steering-lite's loading-weighted Δlogit scale (base log(0.099)
=-2.3, real shift ~0.5-2 nats). foundation_nats now reads rep["profile"].
Also:
- run.py: BLUF `main metric:` line with cue ball (🟢/🟡/🔴 by coherence band).
- heal.py: training table to 2 sig figs (nll/kl/loss .2f, gnorm .1f); a
per-step loss does not warrant 3 decimals.
- diag_stages: accept 1+ ckpts, label each row by its reg from metadata.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The trait metric was taking the diagonal of tinymfv's raw pre-softmax BMA
`score` logit (unnormalised), giving base Authority ~-5 and absurd 8-nat
swings, then comparing those to steering-lite's 0.5-2 nat reference -- which
is a DIFFERENT metric (loading-weighted Delta-logit of binary p(is-wrong)).
Wrong scale, wrong comparison.
Fix: auth_nats = mean log p[authority] on authority-defiance vignettes (the
NORMALIZED choice logprob, the diagonal of the softmax `p`). Base ~log(0.099)
= -2.3, real shifts ~1-3 nats. DRY: evaluate_model now calls foundation_nats.
Also:
- diag_stages: steer at operating point c=0.5 (c=1 collapses coherence to
0.05), add coh_cost = |dCoh|/|dAuth| (coherence lost per nat of behaviour)
to answer "is the adapter a better pareto than raw steering?".
- diag_csweep: drop the bogus 0.5-2 steering-lite anchor; SocialNorms
co-moving with Authority is expected (both binding foundations), not collapse.
- gitignore out/ and results.tsv (experiment outputs, stale schema).
- personas docs (steering-lite proper-pair rules), spec Plans B/C/D, journal.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Root cause found via diag_axis on 4B: raw mean-diff steered across the 7-layer
band (0.4-0.6) at coeff=1 DESTROYS gemma-3-4b (coherence 1.00->0.02). That
starved the filter to 2 kept completions, so the "adapter" was ~untrained
(2 examples) = base behaviour, my Q1 "promising" read was not validated.
Fixes:
- separate steer_layers (narrow 0.45-0.55) for the vector from layer_range
(broad 0.0-1.0) for the LoRA; they were wrongly coupled
- lower alpha sweep (0.25,0.5,1,2); n_prompts=16
- assert len(kept) >= min_train(20); TINY=2. Don't train on starved data.
- heal training table (loguru+tqdm per token-efficient-logging): step, nll, kl,
loss, gnorm + SHOULD
- full untruncated steer + adapter generation dumps with prompt and
coherence(p_ans_any) inline so we can judge coherence/trait ourselves
NOT yet run with fixes on 4B. Base 4B is Care=0.92 (already aligned) -> the
prompting-baseline confound (Q7) is now the critical check.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Per user: no iso-KL calibration. Use raw (unnormalised) mean-diff teacher
vector; sweep cfg.alphas at generation and let the FILTER pick usable C
(filter replaces calibration). Default model google/gemma-3-4b-it (1B too dumb;
Authority degenerate there was a model artifact, not a real conclusion).
Token-efficient discriminating logs so each Q is readable:
- Q0: filter table (alpha -> ppl_mean, kept_frac) + low/high-C samples + SHOULD
- Q1: generate from trained adapter (no steering); adapter_ppl vs steered_ppl
under the original + sample + SHOULD (heal = adapter more coherent than steered)
- Q2/Q3: loop summary table (socialnorms/care/coherence/cos_v0 per round) + SHOULD
fast-dev-run green: ppl rises with alpha (3173->4.2M), adapter_ppl<<steered_ppl.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
scripts/diag_axis.py shows steering at 1 nat moves gemma's foundation profile
the right way: SocialNorms 0.68->0.42, Care 0.21->0.33, coherence 0.72->0.88.
Authority is ~0 on this model (no headroom), so:
- eval reports all foundations; trait axis = SocialNorms (down) + Care (up)
- map.html plots Care vs SocialNorms
- add gen_alpha=1.5: over-steer generation into the incoherent regime so the
heal (Q1) has work to do (at 1 nat coherence improved, nothing to heal)
- results.py groups on coherence/socialnorms/care
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The mean-diff vector is L2-normalised, so p95 KL ~ c^2 and reaching the 1-nat
target needs c ~ O(100). steering-lite's default bracket hi (~16) pinned
c_star at the top (KL ~0.1 << 1.0) on both tiny-random and real gemma. With
bracket=(0.1, 1024) gemma calibrates to c_star=64.03 at p95 KL=1.035.
Also detach div before .item() in heal logging. See RESEARCH_JOURNAL.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Add the by-question results infra per setup-repo conventions:
- results.tsv append at end of each finished run (config + final metrics + argv)
- scripts/results.py groups by arm (reg) into a markdown table; `just results`
- docs/results.md curated by-question snapshot (U2 regulariser comparison)
- docs/{spec,brainstorming,literature,evidence} structure
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Setup per setup-repo conventions: uv + justfile + fast-dev-run on
wassname/qwen3-5lyr-tiny-random, package under src/steer_heal (config +
pipeline skeleton). Stages fail fast with NotImplementedError pointing at
the docs/vendor module to port from.
Design in spec.md: distil a steering-lite mean-diff teacher vector (iso-KL
dosed) into a conditioned LoRA, heal incoherency with a KL-rev-to-original
barrier, fold each round via w2schar gated bake, eval on tinymfv. Three
uncertainty gates (filter / heal / iterate) each with a UAT artifact.
Base model google/gemma-3-1b-it (RTX 3090, 24GB). Reference repos vendored
under docs/vendor (gitignored): steering-lite, isokl, tinymfv, w2schar-mini.
The lighter three are editable path deps; w2schar (py3.13 + flash-attn) is
reference-only, we copy its adapter/bake/plot modules.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
wassname