open-interpretability 0.0.20__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- open_interpretability-0.0.20.dist-info/METADATA +157 -0
- open_interpretability-0.0.20.dist-info/RECORD +256 -0
- open_interpretability-0.0.20.dist-info/WHEEL +4 -0
- open_interpretability-0.0.20.dist-info/entry_points.txt +2 -0
- open_interpretability-0.0.20.dist-info/licenses/LICENSE +21 -0
- openinterp/__init__.py +6 -0
- openinterp/agents/__init__.py +25 -0
- openinterp/agents/base.py +229 -0
- openinterp/agents/judge.py +149 -0
- openinterp/agents/proposer.py +471 -0
- openinterp/agents/proposer_profiles.py +79 -0
- openinterp/agents/review_measurements.py +173 -0
- openinterp/agents/reviewer.py +498 -0
- openinterp/agreement.py +63 -0
- openinterp/audit_options.py +79 -0
- openinterp/baseline.py +254 -0
- openinterp/benchmark.py +450 -0
- openinterp/blends.py +239 -0
- openinterp/builder/__init__.py +6 -0
- openinterp/builder/agents.py +260 -0
- openinterp/builder/aligned.py +41 -0
- openinterp/builder/charts.py +411 -0
- openinterp/builder/circuit_pages.py +37 -0
- openinterp/builder/circuits.py +264 -0
- openinterp/builder/compiler.py +495 -0
- openinterp/builder/corpora.py +98 -0
- openinterp/builder/coverage.py +46 -0
- openinterp/builder/dashboard.py +255 -0
- openinterp/builder/disk_memo.py +65 -0
- openinterp/builder/experiment.py +105 -0
- openinterp/builder/heat.py +121 -0
- openinterp/builder/hrefs.py +71 -0
- openinterp/builder/learned_baseline.py +204 -0
- openinterp/builder/model_facts.py +138 -0
- openinterp/builder/neuronpedia_export.py +107 -0
- openinterp/builder/parallel.py +100 -0
- openinterp/builder/parquet_export.py +458 -0
- openinterp/builder/pieces.py +127 -0
- openinterp/builder/pii.py +211 -0
- openinterp/builder/registry_stats.py +144 -0
- openinterp/builder/research.py +109 -0
- openinterp/builder/sections.py +28 -0
- openinterp/builder/significance.py +103 -0
- openinterp/builder/sister_graph.py +180 -0
- openinterp/builder/standing.py +185 -0
- openinterp/builder/statuses.py +34 -0
- openinterp/builder/stopped.py +91 -0
- openinterp/builder/stops.py +30 -0
- openinterp/builder/superpositions.py +102 -0
- openinterp/builder/templates/_agent_role_cells.html +9 -0
- openinterp/builder/templates/_causal_chart.html +67 -0
- openinterp/builder/templates/_circuit.html +8 -0
- openinterp/builder/templates/_dashboard.html +133 -0
- openinterp/builder/templates/_experiment.html +500 -0
- openinterp/builder/templates/_macros.html +82 -0
- openinterp/builder/templates/_patchscope.html +33 -0
- openinterp/builder/templates/_prior_art.html +24 -0
- openinterp/builder/templates/_specificity.html +21 -0
- openinterp/builder/templates/_steering.html +34 -0
- openinterp/builder/templates/_superposition.html +62 -0
- openinterp/builder/templates/_vocab_chart.html +88 -0
- openinterp/builder/templates/_wiring.html +44 -0
- openinterp/builder/templates/about.html +95 -0
- openinterp/builder/templates/agent.html +69 -0
- openinterp/builder/templates/agents.html +80 -0
- openinterp/builder/templates/base.html +103 -0
- openinterp/builder/templates/circuit.html +64 -0
- openinterp/builder/templates/circuits.html +46 -0
- openinterp/builder/templates/experiment.html +21 -0
- openinterp/builder/templates/hub.html +16 -0
- openinterp/builder/templates/index.html +387 -0
- openinterp/builder/templates/layer.html +69 -0
- openinterp/builder/templates/learned-baseline.html +44 -0
- openinterp/builder/templates/model.html +151 -0
- openinterp/builder/templates/models.html +47 -0
- openinterp/builder/templates/neuron.html +234 -0
- openinterp/builder/templates/paper.html +39 -0
- openinterp/builder/templates/redirect.html +1 -0
- openinterp/builder/templates/research.html +31 -0
- openinterp/builder/templates/user.html +54 -0
- openinterp/builder/templates/users.html +39 -0
- openinterp/builder/templates/wiki/experiment.md +271 -0
- openinterp/builder/templates/wiki/index.md +32 -0
- openinterp/builder/templates/wiki/neuron.md +30 -0
- openinterp/builder/templates/wiki/stop.md +36 -0
- openinterp/builder/tree.py +470 -0
- openinterp/builder/users.py +142 -0
- openinterp/builder/wiki.py +427 -0
- openinterp/checks.py +429 -0
- openinterp/cli.py +494 -0
- openinterp/cli_baseline.py +332 -0
- openinterp/cli_benchmark.py +300 -0
- openinterp/cli_build.py +128 -0
- openinterp/cli_check.py +51 -0
- openinterp/cli_concepts.py +76 -0
- openinterp/cli_corpus.py +252 -0
- openinterp/cli_costs.py +257 -0
- openinterp/cli_coverage.py +76 -0
- openinterp/cli_crossref.py +86 -0
- openinterp/cli_errata.py +142 -0
- openinterp/cli_evidence.py +495 -0
- openinterp/cli_evidence_store.py +93 -0
- openinterp/cli_explain.py +101 -0
- openinterp/cli_loop.py +125 -0
- openinterp/cli_memory.py +28 -0
- openinterp/cli_neuronpedia.py +175 -0
- openinterp/cli_pause.py +45 -0
- openinterp/cli_pole_effects.py +133 -0
- openinterp/cli_probing.py +119 -0
- openinterp/cli_registry.py +348 -0
- openinterp/cli_reproduce.py +446 -0
- openinterp/cli_review.py +302 -0
- openinterp/cli_scan.py +223 -0
- openinterp/cli_score.py +446 -0
- openinterp/cli_sisters.py +86 -0
- openinterp/cli_submission.py +217 -0
- openinterp/cli_superpositions.py +118 -0
- openinterp/cli_wiring.py +54 -0
- openinterp/codeid.py +73 -0
- openinterp/concepts.py +400 -0
- openinterp/core/__init__.py +59 -0
- openinterp/core/agent_names.py +76 -0
- openinterp/core/attribution.py +101 -0
- openinterp/core/cache_dirs.py +15 -0
- openinterp/core/causal.py +82 -0
- openinterp/core/confounds.py +103 -0
- openinterp/core/contributors.py +26 -0
- openinterp/core/corpora.py +126 -0
- openinterp/core/costs.py +430 -0
- openinterp/core/draws.py +109 -0
- openinterp/core/errata.py +244 -0
- openinterp/core/evidence.py +160 -0
- openinterp/core/examples.py +93 -0
- openinterp/core/exceptions.py +29 -0
- openinterp/core/explanation.py +130 -0
- openinterp/core/fence.py +42 -0
- openinterp/core/heavy_lock.py +114 -0
- openinterp/core/identifiers.py +198 -0
- openinterp/core/isolation.py +149 -0
- openinterp/core/ledger.py +130 -0
- openinterp/core/licenses.py +28 -0
- openinterp/core/manifold.py +117 -0
- openinterp/core/partners.py +32 -0
- openinterp/core/patchscope.py +59 -0
- openinterp/core/paths.py +40 -0
- openinterp/core/pause.py +96 -0
- openinterp/core/pricing.py +178 -0
- openinterp/core/prior_art.py +44 -0
- openinterp/core/probing.py +146 -0
- openinterp/core/profiles.py +221 -0
- openinterp/core/project.py +28 -0
- openinterp/core/run_environment.py +38 -0
- openinterp/core/schemas.py +484 -0
- openinterp/core/specificity.py +49 -0
- openinterp/core/stats.py +172 -0
- openinterp/core/steering.py +49 -0
- openinterp/core/strict.py +19 -0
- openinterp/core/takedowns.py +136 -0
- openinterp/core/transcripts.py +130 -0
- openinterp/core/wiring.py +80 -0
- openinterp/corpus.py +476 -0
- openinterp/corpus_cache.py +145 -0
- openinterp/corpus_layouts.py +108 -0
- openinterp/corpus_presets.py +168 -0
- openinterp/corpus_rebuild.py +93 -0
- openinterp/coverage.py +155 -0
- openinterp/crossref.py +442 -0
- openinterp/data/__init__.py +9 -0
- openinterp/data/offensive_words.txt +58 -0
- openinterp/data/public_suffix_list.dat +15771 -0
- openinterp/data/sample_corpus.txt +40 -0
- openinterp/draw_steps.py +359 -0
- openinterp/errata_checks.py +129 -0
- openinterp/evidence.py +462 -0
- openinterp/evidence_branch.py +119 -0
- openinterp/evidence_lenses.py +75 -0
- openinterp/experiment_loop.py +357 -0
- openinterp/explain.py +252 -0
- openinterp/half_precision.py +79 -0
- openinterp/harness/__init__.py +37 -0
- openinterp/harness/ane/__init__.py +9 -0
- openinterp/harness/ane/backend.py +324 -0
- openinterp/harness/ane/cache.py +276 -0
- openinterp/harness/ane/chunks.py +163 -0
- openinterp/harness/ane/convert.py +116 -0
- openinterp/harness/ane/forward.py +258 -0
- openinterp/harness/ane/qualified.py +40 -0
- openinterp/harness/annotations.py +329 -0
- openinterp/harness/attribution.py +96 -0
- openinterp/harness/causal.py +264 -0
- openinterp/harness/confounds.py +184 -0
- openinterp/harness/covariance.py +301 -0
- openinterp/harness/crossref.py +195 -0
- openinterp/harness/cuda_devices.py +88 -0
- openinterp/harness/dead.py +124 -0
- openinterp/harness/devices.py +401 -0
- openinterp/harness/dictionaries.py +496 -0
- openinterp/harness/dreaming.py +116 -0
- openinterp/harness/environment.py +64 -0
- openinterp/harness/explanation_scores.py +371 -0
- openinterp/harness/features.py +276 -0
- openinterp/harness/first_position.py +137 -0
- openinterp/harness/heads.py +226 -0
- openinterp/harness/hooks.py +499 -0
- openinterp/harness/input_attribution.py +367 -0
- openinterp/harness/jlens.py +183 -0
- openinterp/harness/layer_cache.py +301 -0
- openinterp/harness/layer_major.py +274 -0
- openinterp/harness/lookahead.py +93 -0
- openinterp/harness/manifold.py +200 -0
- openinterp/harness/memory.py +471 -0
- openinterp/harness/moe.py +134 -0
- openinterp/harness/patcher.py +322 -0
- openinterp/harness/patchscope.py +177 -0
- openinterp/harness/probing.py +383 -0
- openinterp/harness/scan.py +462 -0
- openinterp/harness/scan_backend.py +201 -0
- openinterp/harness/sister_passes.py +463 -0
- openinterp/harness/sisters.py +363 -0
- openinterp/harness/sites.py +101 -0
- openinterp/harness/specificity.py +107 -0
- openinterp/harness/steering.py +159 -0
- openinterp/harness/superpositions.py +116 -0
- openinterp/harness/units.py +203 -0
- openinterp/harness/wiring.py +432 -0
- openinterp/harness/wiring_activations.py +285 -0
- openinterp/langid.py +73 -0
- openinterp/names.py +129 -0
- openinterp/neuronpedia_baseline.py +88 -0
- openinterp/neuronpedia_review.py +148 -0
- openinterp/planner.py +466 -0
- openinterp/planner_effects.py +89 -0
- openinterp/planner_heads.py +92 -0
- openinterp/planner_models.py +174 -0
- openinterp/planner_sisters.py +122 -0
- openinterp/planner_states.py +57 -0
- openinterp/planner_superpositions.py +114 -0
- openinterp/pole_effects.py +102 -0
- openinterp/prior_art.py +143 -0
- openinterp/reproduce.py +450 -0
- openinterp/revision.py +45 -0
- openinterp/runner_cache.py +35 -0
- openinterp/score_controls.py +169 -0
- openinterp/sensitive.py +243 -0
- openinterp/show.py +236 -0
- openinterp/signing.py +209 -0
- openinterp/sisters.py +476 -0
- openinterp/superposition_finders.py +225 -0
- openinterp/sync/__init__.py +5 -0
- openinterp/sync/r2.py +280 -0
- openinterp/syntax.py +399 -0
- openinterp/takedown_checks.py +76 -0
- openinterp/toxicity.py +158 -0
- openinterp/verdict_regression.py +79 -0
- openinterp/wiring_backfill.py +170 -0
- openinterp/wordnet.py +362 -0
|
@@ -0,0 +1,387 @@
|
|
|
1
|
+
{% extends "base.html" %}
|
|
2
|
+
{% from "_macros.html" import map_fill, status_badge, evidence_badge, self_review_badge, pole_mark, stop_outcome %}
|
|
3
|
+
{% set t = stats.totals %}
|
|
4
|
+
|
|
5
|
+
{% block content %}
|
|
6
|
+
<section class="mb-10">
|
|
7
|
+
<h1 class="text-3xl font-semibold tracking-tight">Neuron registry</h1>
|
|
8
|
+
<p class="mt-2 max-w-3xl text-slate-600 dark:text-slate-400">
|
|
9
|
+
Every experiment on native MLP neuron poles and attention heads in open-weight models, each stress-tested by a
|
|
10
|
+
reviewer agent (ideally a different model from the proposer) with adversarial hard negatives, minimal-pair counterfactuals
|
|
11
|
+
and causal patching. Null results are published too: a well-tested hypothesis that fails
|
|
12
|
+
tells us as much as one that passes.
|
|
13
|
+
</p>
|
|
14
|
+
<dl class="mt-6 grid grid-cols-2 gap-px overflow-hidden rounded-lg border border-slate-200 bg-slate-200 sm:grid-cols-6 dark:border-slate-800 dark:bg-slate-800">
|
|
15
|
+
{% for label, value in [
|
|
16
|
+
("Experiments", t.experiments),
|
|
17
|
+
("Neurons studied", t.neurons),
|
|
18
|
+
("Heads studied", t.heads),
|
|
19
|
+
("Models", t.models),
|
|
20
|
+
("Verified units", t.verified),
|
|
21
|
+
("Conclusive nulls", t.conclusive_nulls),
|
|
22
|
+
] %}
|
|
23
|
+
<div class="bg-white p-4 dark:bg-slate-950">
|
|
24
|
+
<dt class="text-xs uppercase tracking-wide text-slate-500">{{ label }}</dt>
|
|
25
|
+
<dd class="mt-1 font-mono text-2xl tabular-nums">{{ value }}</dd>
|
|
26
|
+
</div>
|
|
27
|
+
{% endfor %}
|
|
28
|
+
</dl>
|
|
29
|
+
<p class="mt-3 text-sm text-slate-600 dark:text-slate-400" title="Sum of (1 − P(useful)) over the current verified claims, one per neuron pole or head">
|
|
30
|
+
{% if fdr.claims %}
|
|
31
|
+
Estimated false discoveries among the {{ fdr.claims }} verified claims:
|
|
32
|
+
<strong class="font-mono">{{ fdr.expected_false }}</strong> (false discovery rate ≈ {{ fdr.rate | pct }}).
|
|
33
|
+
{% else %}
|
|
34
|
+
No verified claims yet, so no false discovery estimate.
|
|
35
|
+
{% endif %}
|
|
36
|
+
Repeat experiments that re-measure the same probes are not counted as independent evidence.
|
|
37
|
+
</p>
|
|
38
|
+
{% if t.features %}
|
|
39
|
+
<p class="mt-1 text-sm text-slate-600 dark:text-slate-400">
|
|
40
|
+
Not counted above: {{ t.features }} latents of sparse dictionaries (transcoders or SAEs) audited under the same protocol as a <a class="underline" href="{{ root }}learned-baseline.html">learned-basis baseline</a>, {{ t.verified_features }} of them verified.
|
|
41
|
+
</p>
|
|
42
|
+
{% endif %}
|
|
43
|
+
{% if t.superpositions %}
|
|
44
|
+
<p class="mt-1 text-sm text-slate-600 dark:text-slate-400">
|
|
45
|
+
Not in the verified units: {{ t.superpositions }} superpositions studied, {{ t.verified_superpositions }} of them verified. A superposition is a sparse direction over the neurons of one layer.
|
|
46
|
+
</p>
|
|
47
|
+
{% endif %}
|
|
48
|
+
{% if t.member_only %}
|
|
49
|
+
<p class="mt-1 text-sm text-slate-600 dark:text-slate-400">
|
|
50
|
+
{{ t.neurons }} neurons studied, {{ t.member_only }} of them only as part of a superposition.
|
|
51
|
+
</p>
|
|
52
|
+
{% endif %}
|
|
53
|
+
{% if stats.protocol_mix %}
|
|
54
|
+
{% set current = stats.protocol_mix | selectattr("protocol_version", "equalto", stats.current_protocol) | map(attribute="units") | sum %}
|
|
55
|
+
<p class="mt-1 text-sm text-slate-600 dark:text-slate-400" title="Each verified unit counts the protocol of its newest record that agrees with its standing">
|
|
56
|
+
Protocol of the verified claims: {{ current }} of {{ t.verified }} were judged under the current protocol ({{ stats.current_protocol }}){% for row in stats.protocol_mix if row.protocol_version != stats.current_protocol %}{{ "; " if loop.first else ", " }}{{ row.units }} under {{ row.protocol_version }}{% endfor %}.
|
|
57
|
+
Older verdicts were judged by older rules (<a class="underline" href="{{ repo_url }}/blob/main/docs/PROTOCOL.md">protocol §6</a>).
|
|
58
|
+
</p>
|
|
59
|
+
{% endif %}
|
|
60
|
+
{% if t.self_reviewed %}
|
|
61
|
+
<p class="mt-1 text-sm text-slate-600 dark:text-slate-400">
|
|
62
|
+
{{ t.self_reviewed }} of {{ t.experiments }} experiments were {{ self_review_badge() }}: the proposer's model also wrote the probes.
|
|
63
|
+
{% if t.verified %}So were {{ t.verified_self_reviewed }} of the {{ t.verified }} current verified claims.{% endif %}
|
|
64
|
+
That is allowed, but a reviewer from a different model is worth more, and <code class="font-mono">open-interpretability next</code> draws re-reviews of these.
|
|
65
|
+
</p>
|
|
66
|
+
{% endif %}
|
|
67
|
+
{% if agreement.claims %}
|
|
68
|
+
<p class="mt-1 text-sm text-slate-600 dark:text-slate-400" title="Each probe set is read by its newest experiment, so re-measurements don't count twice">
|
|
69
|
+
Reviewer agreement: {{ agreement.unanimous }} of the {{ agreement.claims }} hypotheses reviewed with more than one probe set got the same verdict, VERIFIED or not, from every review.
|
|
70
|
+
Pairs of reviews agree {{ agreement.pair_agreement | pct }} of the time{% if agreement.cross_model_pairs %} ({{ agreement.cross_model_agreement | pct }} of {{ agreement.cross_model_pairs }} pairs from different reviewer models){% endif %}.
|
|
71
|
+
{% if contested %}<a class="underline" href="#contested">{{ contested | length }} unit{{ 's are' if contested | length != 1 else ' is' }} contested</a>: a VERIFIED review and another review of its current hypothesis disagree, so it is not counted as verified.{% endif %}
|
|
72
|
+
</p>
|
|
73
|
+
{% endif %}
|
|
74
|
+
</section>
|
|
75
|
+
|
|
76
|
+
{% if dashboard %}{% include "_dashboard.html" %}{% endif %}
|
|
77
|
+
|
|
78
|
+
{% if stats.coverage %}
|
|
79
|
+
<section class="mb-10" id="coverage">
|
|
80
|
+
<h2 class="mb-1 text-lg font-semibold">How much of each model is explained</h2>
|
|
81
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">For each MLP layer, mean-ablating every neuron raises the next-token loss. The share of that rise that comes back when only the VERIFIED poles keep their real activations is the registry's coverage. It is weighted by each layer's rise. The baseline keeps as many random poles. Method: <a class="underline" href="{{ repo_url }}/blob/main/docs/analyses/registry-coverage.md">registry coverage</a>.</p>
|
|
82
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
83
|
+
<table class="w-full text-left text-sm">
|
|
84
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
85
|
+
<tr><th class="px-3 py-2">Model</th><th class="px-3 py-2 text-right">VERIFIED poles</th><th class="px-3 py-2 text-right">Loss effect kept</th><th class="px-3 py-2 text-right">Random poles, same count</th><th class="px-3 py-2 text-right">Measured</th></tr>
|
|
86
|
+
</thead>
|
|
87
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
88
|
+
{% for c in stats.coverage %}
|
|
89
|
+
<tr>
|
|
90
|
+
<td class="px-3 py-2 font-mono text-xs">{{ c.model }}</td>
|
|
91
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ c.verified_poles | thousands }} of {{ c.poles | thousands }}</td>
|
|
92
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ "—" if c.recovered is none else "{:.2%}".format(c.recovered) }}</td>
|
|
93
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ "—" if c.recovered_random is none else "{:.2%}".format(c.recovered_random) }}</td>
|
|
94
|
+
<td class="px-3 py-2 text-right font-mono text-xs">{{ c.date }}</td>
|
|
95
|
+
</tr>
|
|
96
|
+
{% endfor %}
|
|
97
|
+
</tbody>
|
|
98
|
+
</table>
|
|
99
|
+
</div>
|
|
100
|
+
</section>
|
|
101
|
+
{% endif %}
|
|
102
|
+
|
|
103
|
+
{% if stats.baseline %}
|
|
104
|
+
<p class="mb-10 max-w-3xl text-sm text-slate-600 dark:text-slate-400">
|
|
105
|
+
Are a model's own neurons as explainable as the latents of a sparse dictionary?
|
|
106
|
+
{% for b in stats.baseline %}On <span class="font-mono">{{ b.model_name }}</span>, {{ b.features.verified }} of {{ b.features.units }} random latents of <span class="font-mono">{{ b.dictionary }}</span> verified, against {{ b.positive_poles.verified }} of {{ b.positive_poles.units }} random positive poles. {% endfor %}
|
|
107
|
+
The <a class="underline" href="{{ root }}learned-baseline.html">learned-basis baseline</a> page has the tables.
|
|
108
|
+
</p>
|
|
109
|
+
{% endif %}
|
|
110
|
+
|
|
111
|
+
<section class="mb-10">
|
|
112
|
+
<h2 class="mb-3 text-lg font-semibold">Models</h2>
|
|
113
|
+
{% if models %}
|
|
114
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">New neurons are drawn mostly from the smallest model with less than {{ '{:g}%'.format(100 * explore.target) }} of its neurons verified, so that one model is explained well before the next.</p>
|
|
115
|
+
<div class="space-y-4">
|
|
116
|
+
{% for m in models %}
|
|
117
|
+
{% set pc = pieces.model_pieces(m) %}
|
|
118
|
+
<a href="{{ m.href }}" class="block rounded-lg border border-slate-200 p-4 hover:border-emerald-500 dark:border-slate-800">
|
|
119
|
+
<div class="flex flex-wrap items-baseline justify-between gap-2">
|
|
120
|
+
<span class="font-mono text-sm font-medium">{{ m.name }}{% if m.name == explore.model %} <span class="ml-1 rounded bg-emerald-100 px-1.5 py-0.5 font-sans text-xs font-normal text-emerald-800 dark:bg-emerald-900/40 dark:text-emerald-300" title="The smallest model below the verified target: open-interpretability next draws most new neurons from it.">explore focus</span>{% endif %}</span>
|
|
121
|
+
<span class="text-sm tabular-nums text-slate-600 dark:text-slate-400">
|
|
122
|
+
{{ m.experiments }} experiments ·
|
|
123
|
+
{{ m.neurons_studied }}/{{ m.total_neurons | thousands }} neurons studied{% if m.member_only %}, {{ m.member_only }} of them only as part of a superposition{% endif %} ({{ m.coverage | pct_fine }}) ·
|
|
124
|
+
{% set v = explore.models[m.name] %}{{ v.verified }}/{{ m.total_neurons | thousands }} verified ({{ (v.verified / m.total_neurons if m.total_neurons else none) | pct_fine }}; target {{ '{:g}%'.format(100 * explore.target) }}){% if v.dead %}, {{ v.dead }} of them proven dead{% endif %}
|
|
125
|
+
</span>
|
|
126
|
+
</div>
|
|
127
|
+
{% if pc %}
|
|
128
|
+
<svg viewBox="0 0 {{ pc.width }} {{ 4 * pc.rows | length - 1 }}" class="mt-3 w-full" role="img" data-pagefind-ignore
|
|
129
|
+
aria-label="Pieces bars for {{ m.name }}: one row per layer, L0 at the top, each cell {{ pc.block }} neurons; {{ pc.lit }} of {{ pc.cells | thousands }} cells hold a studied neuron.">
|
|
130
|
+
<defs><pattern id="pieces-{{ loop.index }}" width="{{ pc.pitch }}" height="4" patternUnits="userSpaceOnUse"><rect width="{{ pc.pitch - 1 }}" height="3" class="fill-slate-200 dark:fill-slate-800"/></pattern></defs>
|
|
131
|
+
<rect width="{{ pc.width + 1 }}" height="{{ 4 * pc.rows | length }}" fill="url(#pieces-{{ loop.index }})"/>
|
|
132
|
+
{% for row in pc.rows %}
|
|
133
|
+
{% for c in row.cells %}
|
|
134
|
+
<rect x="{{ c.x }}" y="{{ 4 * row.layer }}" width="{{ pc.pitch - 1 }}" height="3" class="{{ map_fill[c.status] }}"><title>{{ c.label }}</title></rect>
|
|
135
|
+
{% endfor %}
|
|
136
|
+
{% endfor %}
|
|
137
|
+
</svg>
|
|
138
|
+
<p class="mt-1 text-xs text-slate-500">One row per layer, L0 at the top; each cell is {{ pc.block }} neurons, lit when any is studied and coloured by the best status among them.</p>
|
|
139
|
+
{% endif %}
|
|
140
|
+
</a>
|
|
141
|
+
{% endfor %}
|
|
142
|
+
</div>
|
|
143
|
+
{% else %}
|
|
144
|
+
<p class="text-sm text-slate-500">No experiments yet.</p>
|
|
145
|
+
{% endif %}
|
|
146
|
+
</section>
|
|
147
|
+
|
|
148
|
+
{% if pageless_stops %}
|
|
149
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
150
|
+
<h2 class="mb-1 text-lg font-semibold">Scans that stopped on units with no page</h2>
|
|
151
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">A scan stops on a unit that never fires or that is active on at least half of its tokens. No experiment exists on these units, so they have no page here. These attempts tested no hypothesis, so they give no verdict.</p>
|
|
152
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
153
|
+
<table class="w-full text-left text-sm">
|
|
154
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
155
|
+
<tr><th class="px-3 py-2">Date</th><th class="px-3 py-2">Unit</th><th class="px-3 py-2">Outcome</th><th class="px-3 py-2 text-right">Active share</th><th class="px-3 py-2">Page</th></tr>
|
|
156
|
+
</thead>
|
|
157
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
158
|
+
{% for v in pageless_stops %}
|
|
159
|
+
<tr class="align-top">
|
|
160
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono text-xs"><time datetime="{{ v.stop.timestamp.isoformat() }}">{{ v.stop.timestamp.strftime("%Y-%m-%d %H:%M") }}</time></td>
|
|
161
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono text-xs">{{ v.label }}<div class="text-slate-500">{{ v.stop.identifier.model_name }}</div></td>
|
|
162
|
+
<td class="px-3 py-2">{{ stop_outcome(v.stop) }}</td>
|
|
163
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ v.stop.active_share | pct }}</td>
|
|
164
|
+
<td class="px-3 py-2 text-xs"><a class="underline-offset-2 hover:underline" href="{{ repo_url }}/blob/main/{{ v.page }}">Wiki page</a></td>
|
|
165
|
+
</tr>
|
|
166
|
+
{% endfor %}
|
|
167
|
+
</tbody>
|
|
168
|
+
</table>
|
|
169
|
+
</div>
|
|
170
|
+
</section>
|
|
171
|
+
{% endif %}
|
|
172
|
+
|
|
173
|
+
<section class="mb-10">
|
|
174
|
+
<h2 class="mb-1 text-lg font-semibold">Most recent experiments</h2>
|
|
175
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">The latest {{ stats.recent | length }} experiments, whatever their verdict, newest first.</p>
|
|
176
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
177
|
+
<table class="w-full text-left text-sm">
|
|
178
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
179
|
+
<tr><th class="hidden px-3 py-2 sm:table-cell">Reviewed</th><th class="px-3 py-2">Unit</th><th class="px-3 py-2">Result</th><th class="hidden px-3 py-2 md:table-cell">Agents</th></tr>
|
|
180
|
+
</thead>
|
|
181
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
182
|
+
{% for r in stats.recent %}
|
|
183
|
+
<tr class="align-top hover:bg-slate-50 dark:hover:bg-slate-900">
|
|
184
|
+
<td class="hidden whitespace-nowrap px-3 py-2 font-mono text-xs text-slate-600 sm:table-cell dark:text-slate-400">
|
|
185
|
+
<time datetime="{{ r.reviewed_iso }}">{{ r.reviewed_date }}</time>
|
|
186
|
+
<div class="text-slate-500">{{ r.reviewed_time }}</div>
|
|
187
|
+
</td>
|
|
188
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono text-xs">
|
|
189
|
+
<a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ r.href }}">L{{ r.layer }}/{{ r.unit }}{% if r.pole %}{{ pole_mark(r.pole) }}{% endif %}</a>
|
|
190
|
+
<div class="text-slate-500">{{ r.model_name }}</div>
|
|
191
|
+
<div class="text-slate-500 sm:hidden">{{ r.reviewed_date }}</div>
|
|
192
|
+
</td>
|
|
193
|
+
<td class="px-3 py-2">
|
|
194
|
+
<div class="mb-1 flex flex-wrap gap-1">{{ status_badge(r.status) }} {{ evidence_badge(r.conclusiveness, r.evidence_direction) }}
|
|
195
|
+
{% if r.self_reviewed %}{{ self_review_badge() }}{% endif %}
|
|
196
|
+
{% if r.remeasurement %}<span class="rounded bg-slate-100 px-2 py-0.5 text-xs text-slate-600 dark:bg-slate-800 dark:text-slate-400" title="Re-measures an earlier experiment's probes; not independent evidence">re-measurement</span>{% endif %}
|
|
197
|
+
</div>
|
|
198
|
+
{{ r.explanation | truncate(140) }}
|
|
199
|
+
</td>
|
|
200
|
+
<td class="hidden px-3 py-2 font-mono text-xs text-slate-500 md:table-cell">
|
|
201
|
+
<div title="proposer"><a class="underline-offset-2 hover:underline" href="{{ root }}{{ agent_href(r.submitting_agent_model) }}" title="recorded as {{ r.submitting_agent_model }}">{{ agent_name(r.submitting_agent_model) }}</a></div>
|
|
202
|
+
<div title="reviewer">↳ <a class="underline-offset-2 hover:underline" href="{{ root }}{{ agent_href(r.reviewer_agent_model) }}" title="recorded as {{ r.reviewer_agent_model }}">{{ agent_name(r.reviewer_agent_model) }}</a></div>
|
|
203
|
+
</td>
|
|
204
|
+
</tr>
|
|
205
|
+
{% else %}
|
|
206
|
+
<tr><td class="px-3 py-2 text-slate-500" colspan="4">No experiments yet.</td></tr>
|
|
207
|
+
{% endfor %}
|
|
208
|
+
</tbody>
|
|
209
|
+
</table>
|
|
210
|
+
</div>
|
|
211
|
+
</section>
|
|
212
|
+
|
|
213
|
+
<section class="mb-10">
|
|
214
|
+
<h2 class="mb-1 text-lg font-semibold">Most conclusive results</h2>
|
|
215
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">Strongest evidence in either direction, from every experiment.</p>
|
|
216
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
217
|
+
<table class="w-full text-left text-sm">
|
|
218
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
219
|
+
{% for r in stats.conclusive %}
|
|
220
|
+
<tr class="align-top hover:bg-slate-50 dark:hover:bg-slate-900">
|
|
221
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono text-xs">
|
|
222
|
+
<a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ r.href }}">L{{ r.layer }}/{{ r.unit }}{% if r.pole %}{{ pole_mark(r.pole) }}{% endif %}</a>
|
|
223
|
+
<div class="text-slate-500">{{ r.model_name }}</div>
|
|
224
|
+
</td>
|
|
225
|
+
<td class="px-3 py-2">
|
|
226
|
+
<div class="mb-1 flex flex-wrap gap-1">{{ evidence_badge(r.conclusiveness, r.evidence_direction) }} {{ status_badge(r.status) }}</div>
|
|
227
|
+
{{ r.explanation | truncate(140) }}
|
|
228
|
+
</td>
|
|
229
|
+
</tr>
|
|
230
|
+
{% else %}
|
|
231
|
+
<tr><td class="px-3 py-2 text-slate-500">No experiments yet.</td></tr>
|
|
232
|
+
{% endfor %}
|
|
233
|
+
</tbody>
|
|
234
|
+
</table>
|
|
235
|
+
</div>
|
|
236
|
+
</section>
|
|
237
|
+
|
|
238
|
+
<p class="mb-10 text-sm text-slate-600 dark:text-slate-400">
|
|
239
|
+
{{ agents | length }} agent model{{ "" if agents | length == 1 else "s" }} proposed, reviewed or judged these experiments{% if agents %}, most often {% for a in agents[:3] %}<a class="font-mono text-xs underline" href="{{ root }}{{ a.href }}">{{ a.model }}</a>{{ ", " if not loop.last }}{% endfor %}{% endif %}.
|
|
240
|
+
<a class="underline" href="{{ root }}agents/index.html">Agents</a> has each one's VERIFIED rate and costs by role.
|
|
241
|
+
</p>
|
|
242
|
+
|
|
243
|
+
{% if contested %}
|
|
244
|
+
<section id="contested" class="mb-10 scroll-mt-4">
|
|
245
|
+
<h2 class="mb-1 text-lg font-semibold">Contested neurons and heads</h2>
|
|
246
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">Reviews of the current hypothesis disagree. Independent reviews outweigh self-reviews, so these are disagreements among reviews of the same kind. Another independent review, or a narrower revision, settles them.</p>
|
|
247
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
248
|
+
<table class="w-full text-left text-sm">
|
|
249
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
250
|
+
<tr><th class="px-3 py-2">Unit</th><th class="px-3 py-2">Hypothesis</th><th class="px-3 py-2">Reviews</th></tr>
|
|
251
|
+
</thead>
|
|
252
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
253
|
+
{% for c in contested %}
|
|
254
|
+
<tr class="align-top hover:bg-slate-50 dark:hover:bg-slate-900">
|
|
255
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono text-xs">
|
|
256
|
+
<a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ c.unit.href }}">L{{ c.unit.layer }}/{{ c.unit.label }}{% if c.unit.is_neuron %}{{ pole_mark(c.side) }}{% endif %}</a>
|
|
257
|
+
<div class="text-slate-500">{{ c.model }}</div>
|
|
258
|
+
</td>
|
|
259
|
+
<td class="px-3 py-2">{% if c.standing.self_reviewed %}<div class="mb-1">{{ self_review_badge() }}</div>{% endif %}{{ c.standing.lead.record.submission.natural_language_explanation | truncate(140) }}</td>
|
|
260
|
+
<td class="space-y-1 px-3 py-2 text-xs">
|
|
261
|
+
{% for r in c.standing.readings %}
|
|
262
|
+
<div class="flex flex-wrap items-center gap-2"><a href="{{ r.href }}">{{ status_badge(r.record.verdict.status) }}</a><a class="font-mono text-slate-500 underline-offset-2 hover:underline" href="{{ root }}{{ agent_href(r.record.verdict.reviewer_agent_model) }}" title="recorded as {{ r.record.verdict.reviewer_agent_model }}">{{ agent_name(r.record.verdict.reviewer_agent_model) }}</a></div>
|
|
263
|
+
{% endfor %}
|
|
264
|
+
</td>
|
|
265
|
+
</tr>
|
|
266
|
+
{% endfor %}
|
|
267
|
+
</tbody>
|
|
268
|
+
</table>
|
|
269
|
+
</div>
|
|
270
|
+
</section>
|
|
271
|
+
{% endif %}
|
|
272
|
+
|
|
273
|
+
{% if corpora %}
|
|
274
|
+
<section class="mb-10" id="corpora">
|
|
275
|
+
<h2 class="mb-1 text-lg font-semibold">Corpora</h2>
|
|
276
|
+
<p class="mb-3 max-w-3xl text-sm text-slate-600 dark:text-slate-400">
|
|
277
|
+
The corpora experiments were scanned on. Registry statistics over randomly selected units describe a model on its corpus, so they depend on what the corpus holds.
|
|
278
|
+
Where documents carry their language, the mix is each model's scanned tokens by language, from its largest scan of the corpus; a tokenizer splits some languages into more tokens than others.
|
|
279
|
+
</p>
|
|
280
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
281
|
+
<table class="w-full text-left text-sm">
|
|
282
|
+
<thead class="text-xs uppercase tracking-wide text-slate-500">
|
|
283
|
+
<tr><th class="px-3 py-2">Corpus</th><th class="px-3 py-2 text-right">Experiments</th><th class="px-3 py-2">Language mix</th></tr>
|
|
284
|
+
</thead>
|
|
285
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
286
|
+
{% for c in corpora %}
|
|
287
|
+
{% set src = corpus_sources.get(c.sha256) %}
|
|
288
|
+
<tr class="align-top">
|
|
289
|
+
<td class="px-3 py-2"><span class="text-xs">{{ src or c.name }}</span> <span class="font-mono text-xs text-slate-500" title="{{ c.sha256 }}">{{ (c.sha256 or "?")[:8] }}</span></td>
|
|
290
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ c.experiments }}</td>
|
|
291
|
+
<td class="px-3 py-2 text-xs">
|
|
292
|
+
{% for mix in c.mixes %}
|
|
293
|
+
<div><span class="font-mono">{{ mix.model_name }}</span>{% if mix.sampled %} <span class="text-slate-500" title="From a --max-tokens sample of the corpus">(sample)</span>{% endif %}: <span class="font-mono" title="{{ "{:,}".format(mix.tokens) }} labelled tokens">{% for l in mix.shown %}{{ l.label }} {{ l.share | pct }}{{ " · " if not loop.last }}{% endfor %}{% if mix.other >= 0.005 %} · other {{ mix.other | pct }}{% endif %}</span></div>
|
|
294
|
+
{% else %}
|
|
295
|
+
<span class="text-slate-500">not labelled by language</span>
|
|
296
|
+
{% endfor %}
|
|
297
|
+
</td>
|
|
298
|
+
</tr>
|
|
299
|
+
{% endfor %}
|
|
300
|
+
</tbody>
|
|
301
|
+
</table>
|
|
302
|
+
</div>
|
|
303
|
+
</section>
|
|
304
|
+
{% endif %}
|
|
305
|
+
|
|
306
|
+
<section>
|
|
307
|
+
<h2 class="mb-1 text-lg font-semibold">Query the registry</h2>
|
|
308
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">
|
|
309
|
+
Runs DuckDB-Wasm in your browser against <code class="font-mono">data/registry.parquet</code>, exposed as the view <code class="font-mono">registry</code>.
|
|
310
|
+
</p>
|
|
311
|
+
<textarea id="sql" rows="3" spellcheck="false"
|
|
312
|
+
class="w-full rounded-lg border border-slate-300 bg-slate-50 p-3 font-mono text-sm dark:border-slate-700 dark:bg-slate-900">SELECT model_name, layer, neuron_idx, pole, head_idx, status, round(conclusiveness, 2) AS conclusive, evidence_direction, explanation
|
|
313
|
+
FROM registry ORDER BY conclusiveness DESC LIMIT 20</textarea>
|
|
314
|
+
<div class="mt-2 flex items-center gap-3">
|
|
315
|
+
<button id="run-sql" class="rounded-md bg-slate-900 px-3 py-1.5 text-sm font-medium text-white hover:bg-slate-700 dark:bg-slate-100 dark:text-slate-900 dark:hover:bg-slate-300">Run query</button>
|
|
316
|
+
<span id="sql-status" class="text-xs text-slate-500"></span>
|
|
317
|
+
</div>
|
|
318
|
+
<div id="sql-result" class="mt-3 overflow-x-auto"></div>
|
|
319
|
+
</section>
|
|
320
|
+
{% endblock %}
|
|
321
|
+
|
|
322
|
+
{% block scripts %}
|
|
323
|
+
<script type="module">
|
|
324
|
+
const DUCKDB = "https://cdn.jsdelivr.net/npm/@duckdb/duckdb-wasm@1.29.0/+esm";
|
|
325
|
+
const status = document.getElementById("sql-status");
|
|
326
|
+
const out = document.getElementById("sql-result");
|
|
327
|
+
let conn;
|
|
328
|
+
|
|
329
|
+
async function connect() {
|
|
330
|
+
const duckdb = await import(DUCKDB);
|
|
331
|
+
const bundle = await duckdb.selectBundle(duckdb.getJsDelivrBundles());
|
|
332
|
+
const workerUrl = URL.createObjectURL(
|
|
333
|
+
new Blob([`importScripts("${bundle.mainWorker}");`], { type: "text/javascript" })
|
|
334
|
+
);
|
|
335
|
+
const db = new duckdb.AsyncDuckDB(new duckdb.VoidLogger(), new Worker(workerUrl));
|
|
336
|
+
await db.instantiate(bundle.mainModule, bundle.pthreadWorker);
|
|
337
|
+
URL.revokeObjectURL(workerUrl);
|
|
338
|
+
const url = new URL("data/registry.parquet", window.location.href).href;
|
|
339
|
+
await db.registerFileURL("registry.parquet", url, duckdb.DuckDBDataProtocol.HTTP, false);
|
|
340
|
+
const conn = await db.connect();
|
|
341
|
+
await conn.query("CREATE VIEW registry AS SELECT * FROM read_parquet('registry.parquet')");
|
|
342
|
+
return conn;
|
|
343
|
+
}
|
|
344
|
+
|
|
345
|
+
function cell(value) {
|
|
346
|
+
const td = document.createElement("td");
|
|
347
|
+
td.className = "px-3 py-1.5 align-top font-mono text-xs";
|
|
348
|
+
td.textContent = value === null || value === undefined ? "∅"
|
|
349
|
+
: typeof value === "object" ? JSON.stringify(value, (_k, v) => typeof v === "bigint" ? v.toString() : v)
|
|
350
|
+
: String(value);
|
|
351
|
+
return td;
|
|
352
|
+
}
|
|
353
|
+
|
|
354
|
+
function render(table) {
|
|
355
|
+
const fields = table.schema.fields.map((f) => f.name);
|
|
356
|
+
const el = document.createElement("table");
|
|
357
|
+
el.className = "w-full rounded-lg border border-slate-200 text-left dark:border-slate-800";
|
|
358
|
+
const head = el.createTHead().insertRow();
|
|
359
|
+
for (const name of fields) {
|
|
360
|
+
const th = document.createElement("th");
|
|
361
|
+
th.className = "bg-slate-50 px-3 py-1.5 text-xs font-medium dark:bg-slate-900";
|
|
362
|
+
th.textContent = name;
|
|
363
|
+
head.appendChild(th);
|
|
364
|
+
}
|
|
365
|
+
const body = el.createTBody();
|
|
366
|
+
for (const row of table.toArray()) {
|
|
367
|
+
const tr = body.insertRow();
|
|
368
|
+
tr.className = "border-t border-slate-200 dark:border-slate-800";
|
|
369
|
+
const obj = row.toJSON();
|
|
370
|
+
for (const name of fields) tr.appendChild(cell(obj[name]));
|
|
371
|
+
}
|
|
372
|
+
out.replaceChildren(el);
|
|
373
|
+
status.textContent = `${table.numRows} rows`;
|
|
374
|
+
}
|
|
375
|
+
|
|
376
|
+
document.getElementById("run-sql").addEventListener("click", async () => {
|
|
377
|
+
status.textContent = conn ? "Running…" : "Loading DuckDB-Wasm…";
|
|
378
|
+
try {
|
|
379
|
+
conn ??= await connect();
|
|
380
|
+
render(await conn.query(document.getElementById("sql").value));
|
|
381
|
+
} catch (err) {
|
|
382
|
+
status.textContent = "";
|
|
383
|
+
out.textContent = String(err);
|
|
384
|
+
}
|
|
385
|
+
});
|
|
386
|
+
</script>
|
|
387
|
+
{% endblock %}
|
|
@@ -0,0 +1,69 @@
|
|
|
1
|
+
{% extends "base.html" %}
|
|
2
|
+
{% from "_macros.html" import breadcrumb, map_fill, status_badge, evidence_badge, self_review_badge, pole_mark %}
|
|
3
|
+
|
|
4
|
+
{% block title %}{{ model.name }} layer {{ layer.layer }} · {{ site_title }}{% endblock %}
|
|
5
|
+
|
|
6
|
+
{% block content %}
|
|
7
|
+
{{ breadcrumb(root, model, layer) }}
|
|
8
|
+
<h1 class="text-2xl font-semibold"><span class="font-mono">{{ model.name }}</span> · layer {{ layer.layer }}</h1>
|
|
9
|
+
<p class="mt-2 text-sm text-slate-600 dark:text-slate-400">
|
|
10
|
+
{{ layer.neurons_studied }} neurons studied{% if layer.member_only %}, {{ layer.member_only }} of them only as part of a superposition{% endif %}{% if layer.stopped %} ({{ layer.stopped | length }} with only stopped scans){% endif %} · {% if layer.heads %}{{ layer.heads | length }} heads studied · {% endif %}{{ layer.experiments }} experiments · {{ layer.verified }} with a verified pole{% if layer.heads %} · {{ layer.verified_heads }} verified heads{% endif %}{% if layer.features %} · {{ layer.features | length }} dictionary features studied as a baseline ({{ layer.verified_features }} verified){% endif %}{% if layer.superpositions %} · {{ layer.superpositions | length }} superpositions studied ({{ layer.verified_superpositions }} verified){% endif %}
|
|
11
|
+
</p>
|
|
12
|
+
|
|
13
|
+
{% set lp = pieces.layer_pieces(model, layer) %}
|
|
14
|
+
{% if lp %}
|
|
15
|
+
<figure class="mt-6" data-pagefind-ignore>
|
|
16
|
+
<div class="overflow-x-auto">
|
|
17
|
+
<svg viewBox="0 0 {{ lp.width }} {{ lp.height }}" class="w-full min-w-[520px]" role="img" aria-label="Pieces map of layer {{ layer.layer }}: one cell per neuron, {{ lp.rows[0].width // lp.pitch + 1 }} to a row, {{ layer.studied }} of {{ model.mlp_width | thousands }} studied ({% for s, n in lp.counts.items() %}{{ n }} {{ s | replace('_', ' ') }}{{ ', ' if not loop.last }}{% endfor %}). Each studied neuron is in the table below.">
|
|
18
|
+
<defs><pattern id="pieces" width="{{ lp.pitch }}" height="{{ lp.pitch }}" patternUnits="userSpaceOnUse"><rect width="{{ lp.pitch - 1 }}" height="{{ lp.pitch - 1 }}" class="fill-slate-100 dark:fill-slate-800/60"/></pattern></defs>
|
|
19
|
+
{% for r in lp.rows %}
|
|
20
|
+
<rect x="{{ lp.left }}" y="{{ r.y }}" width="{{ r.width + 1 }}" height="{{ lp.pitch }}" fill="url(#pieces)"/>
|
|
21
|
+
{% if loop.index0 % lp.label_every == 0 %}<text x="{{ lp.left - 4 }}" y="{{ r.y + 7 }}" text-anchor="end" class="fill-slate-500 font-mono" font-size="8">N{{ r.start | thousands }}</text>{% endif %}
|
|
22
|
+
{% endfor %}
|
|
23
|
+
{% for c in lp.cells %}
|
|
24
|
+
{% if c.href %}<a href="{{ root }}{{ c.href }}">{% endif %}<title>{{ c.label }}</title><rect x="{{ c.x }}" y="{{ c.y }}" width="{{ lp.pitch - 1 }}" height="{{ lp.pitch - 1 }}" class="{{ map_fill[c.status] }}"/>{% if c.href %}</a>{% endif %}
|
|
25
|
+
{% endfor %}
|
|
26
|
+
</svg>
|
|
27
|
+
</div>
|
|
28
|
+
<figcaption class="mt-1 flex flex-wrap items-center gap-x-4 gap-y-1 text-xs text-slate-500">
|
|
29
|
+
{% for s, n in lp.counts.items() %}
|
|
30
|
+
<span><svg class="inline align-middle" width="12" height="12" aria-hidden="true"><rect x="1" y="1" width="10" height="10" rx="2" class="{{ map_fill[s] }}"/></svg> {{ s | replace('_', ' ') }} {{ n }}</span>
|
|
31
|
+
{% endfor %}
|
|
32
|
+
<span><svg class="inline align-middle" width="12" height="12" aria-hidden="true"><rect x="1" y="1" width="10" height="10" rx="2" class="fill-slate-100 dark:fill-slate-800/60"/></svg> unexplored {{ lp.unexplored | thousands }}</span>
|
|
33
|
+
<span>Each cell is one neuron, in index order left to right; a studied one is coloured by its best pole's status and links to its page. STOPPED: every scan of the neuron stopped because it was dead or default-on, so it has no experiment or page.</span>
|
|
34
|
+
</figcaption>
|
|
35
|
+
</figure>
|
|
36
|
+
{% endif %}
|
|
37
|
+
|
|
38
|
+
<div class="mt-6 overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
39
|
+
<table class="w-full text-left text-sm">
|
|
40
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
41
|
+
<tr>
|
|
42
|
+
<th class="px-3 py-2">Unit</th>
|
|
43
|
+
<th class="px-3 py-2">Latest status by pole</th>
|
|
44
|
+
<th class="px-3 py-2 text-right">Experiments</th>
|
|
45
|
+
<th class="px-3 py-2">Most conclusive result</th>
|
|
46
|
+
</tr>
|
|
47
|
+
</thead>
|
|
48
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
49
|
+
{% for n in layer.units %}
|
|
50
|
+
<tr class="align-top hover:bg-slate-50 dark:hover:bg-slate-900">
|
|
51
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono">
|
|
52
|
+
<a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ root }}{{ n.href }}">{{ n.label }}</a>{% if n.dictionary %}<div class="text-xs text-slate-500">{{ n.dictionary }}</div>{% endif +%}
|
|
53
|
+
</td>
|
|
54
|
+
<td class="space-y-1 px-3 py-2">
|
|
55
|
+
{% for pole, s in n.standings | dictsort(reverse=true) %}
|
|
56
|
+
<div class="flex items-center gap-2 text-xs">{% if n.is_neuron %}{{ pole_mark(pole) }} {% endif %}{{ status_badge(s.status) }}{% if s.self_reviewed %} {{ self_review_badge() }}{% endif %}</div>
|
|
57
|
+
{% endfor %}
|
|
58
|
+
</td>
|
|
59
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ n.experiments | length }}</td>
|
|
60
|
+
<td class="px-3 py-2">
|
|
61
|
+
<div class="mb-1">{{ evidence_badge(n.top.sig.conclusiveness, n.top.sig.direction) }}</div>
|
|
62
|
+
<a class="hover:underline" href="{{ root }}{{ n.top.href }}">{{ n.top.record.submission.natural_language_explanation | truncate(140) }}</a>
|
|
63
|
+
</td>
|
|
64
|
+
</tr>
|
|
65
|
+
{% endfor %}
|
|
66
|
+
</tbody>
|
|
67
|
+
</table>
|
|
68
|
+
</div>
|
|
69
|
+
{% endblock %}
|
|
@@ -0,0 +1,44 @@
|
|
|
1
|
+
{% extends "base.html" %}
|
|
2
|
+
|
|
3
|
+
{% block title %}Learned-basis baseline · {{ site_title }}{% endblock %}
|
|
4
|
+
|
|
5
|
+
{% block content %}
|
|
6
|
+
<nav class="mb-4 text-xs text-slate-500"><a class="underline" href="{{ root }}index.html">Registry</a> / Learned-basis baseline</nav>
|
|
7
|
+
<h1 class="mb-4 text-2xl font-semibold">Learned-basis baseline</h1>
|
|
8
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">The paper <a class="underline" href="{{ root }}research/native-vs-learned/index.html">Native neurons against learned dictionaries, at equal effort</a> discusses these numbers, per unit and per layer.</p>
|
|
9
|
+
<section class="mb-10" id="learned-baseline">
|
|
10
|
+
|
|
11
|
+
<p class="mb-3 max-w-3xl text-sm text-slate-600 dark:text-slate-400">
|
|
12
|
+
Are a model's own neurons as explainable as the latents of a sparse dictionary trained on its MLPs?
|
|
13
|
+
Randomly selected latents and randomly selected neuron poles of the same model and layers, scanned the same way and audited under the same protocol.
|
|
14
|
+
Each unit counts once, by its first experiment, so follow-ups don't favour either side. Intervals are 95% Wilson intervals;
|
|
15
|
+
p is a two-sided Fisher exact test of latents against positive poles (and against all poles). In a GELU MLP, such as GPT-2's, the negative pole sits in the activation's dip on most tokens, so no hard negative can pass; there the positive poles are the comparison.
|
|
16
|
+
</p>
|
|
17
|
+
{% for b in stats.baseline %}
|
|
18
|
+
<div class="mb-4 overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
19
|
+
<table class="w-full text-left text-sm">
|
|
20
|
+
<caption class="bg-slate-50 px-3 py-2 text-left text-xs text-slate-600 dark:bg-slate-900 dark:text-slate-400">
|
|
21
|
+
<span class="font-mono">{{ b.model_name }}</span> · <span class="font-mono">{{ b.dictionary }}</span> · layers {{ b.layers | first }}–{{ b.layers | last }} ·
|
|
22
|
+
{% if b.window_tokens %}{{ b.window_tokens }}-token windows{% else %}whole documents{% endif %} of corpus <span class="font-mono" title="{{ b.corpus_sha256 }}">{{ (b.corpus_sha256 or "?")[:8] }}</span> ·
|
|
23
|
+
explained variance {{ b.explained_variance | pct }} · mean L0 {{ b.mean_l0 | num(0) }}{% if b.dead_share is not none %} · <span title="Latents inactive on every scanned token: a random draw of one can't be audited, so they are not in the latents' row">dead latents {{ b.dead_share | pct }}</span>{% endif %}{% if b.p_positive is not none %} · p = {{ b.p_positive | pval }} vs positive poles{% endif %}{% if b.p_value is not none %} ({{ b.p_value | pval }} vs all){% endif %}
|
|
24
|
+
</caption>
|
|
25
|
+
<thead class="text-xs uppercase tracking-wide text-slate-500">
|
|
26
|
+
<tr><th class="px-3 py-2">Random units</th><th class="px-3 py-2 text-right">Audited</th><th class="px-3 py-2 text-right">Verified</th><th class="px-3 py-2 text-right">Share (95% CI)</th><th class="px-3 py-2 text-right">Balanced accuracy</th><th class="px-3 py-2 text-right" title="Random draws the scan stopped on, recorded in wiki/draws/: never active (dead), or active on at least half the scan's tokens (default-on)">Not auditable</th></tr>
|
|
27
|
+
</thead>
|
|
28
|
+
<tbody class="divide-y divide-slate-200 tabular-nums dark:divide-slate-800">
|
|
29
|
+
{% for label, arm in [("Dictionary latents", b.features), ("Neuron poles", b.neurons), ("↳ positive poles", b.positive_poles), ("↳ negative poles", b.negative_poles)] %}
|
|
30
|
+
<tr>
|
|
31
|
+
<td class="px-3 py-2">{{ label }}</td>
|
|
32
|
+
<td class="px-3 py-2 text-right font-mono">{{ arm.units }}</td>
|
|
33
|
+
<td class="px-3 py-2 text-right font-mono">{{ arm.verified }}</td>
|
|
34
|
+
<td class="px-3 py-2 text-right font-mono">{% if arm.units %}{{ arm.share | pct }} ({{ arm.interval[0] | pct }}–{{ arm.interval[1] | pct }}){% else %}—{% endif %}</td>
|
|
35
|
+
<td class="px-3 py-2 text-right font-mono">{{ arm.mean_balanced_accuracy | num(2) }}</td>
|
|
36
|
+
<td class="px-3 py-2 text-right font-mono">{% if arm.dead or arm.default_on %}{% if arm.dead %}{{ arm.dead }} dead{% endif %}{% if arm.dead and arm.default_on %}, {% endif %}{% if arm.default_on %}{{ arm.default_on }} default-on{% endif %}{% else %}—{% endif %}</td>
|
|
37
|
+
</tr>
|
|
38
|
+
{% endfor %}
|
|
39
|
+
</tbody>
|
|
40
|
+
</table>
|
|
41
|
+
</div>
|
|
42
|
+
{% endfor %}
|
|
43
|
+
</section>
|
|
44
|
+
{% endblock %}
|