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 +1,33 @@
|
|
|
1
|
+
{# Patchscope readout for _experiment.html. Expects `sub`. #}
|
|
2
|
+
{% if sub.patchscope and sub.patchscope.readings %}
|
|
3
|
+
{% set ps = sub.patchscope %}
|
|
4
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
5
|
+
<h3 class="mb-1 text-base font-semibold" title="Ghandeharioun et al., 2024; Chen et al., 2024 (SelfIE)">Patchscope</h3>
|
|
6
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">The model's own description of this unit's output direction, with no examples. The direction replaces the residual stream after the named layer at the <code class="tok">{{ ps.placeholder }}</code> of each prompt, at each multiple of the norm it replaces, and the model continues greedily. The control injects a random direction instead, so text the two share comes from the prompt.</p>
|
|
7
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
8
|
+
<table class="w-full text-left text-xs">
|
|
9
|
+
<thead class="bg-slate-50 uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
10
|
+
<tr>
|
|
11
|
+
<th class="px-3 py-2">Prompt</th><th class="px-3 py-2">Layer</th><th class="px-3 py-2"></th>
|
|
12
|
+
{% for s in ps.scales %}<th class="px-3 py-2" title="Norm of the injected vector over the norm of the residual it replaces">×{{ "%g" | format(s) }}</th>{% endfor %}
|
|
13
|
+
</tr>
|
|
14
|
+
</thead>
|
|
15
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
16
|
+
{% for prompt, layer, readings in ps.rows %}
|
|
17
|
+
<tr>
|
|
18
|
+
<td class="px-3 py-1 align-top" rowspan="2" title="{{ ps.prompts[prompt] | replace('{x}', ps.placeholder) }}">{{ prompt }}</td>
|
|
19
|
+
<td class="px-3 py-1 align-top font-mono tabular-nums" rowspan="2">{{ layer }}</td>
|
|
20
|
+
<td class="px-3 py-1 text-slate-500">unit</td>
|
|
21
|
+
{% for r in readings %}<td class="px-3 py-1 font-mono">{{ r.text }}</td>{% endfor %}
|
|
22
|
+
</tr>
|
|
23
|
+
<tr class="text-slate-500">
|
|
24
|
+
<td class="px-3 py-1">control</td>
|
|
25
|
+
{% for r in readings %}<td class="px-3 py-1 font-mono">{{ r.control_text }}</td>{% endfor %}
|
|
26
|
+
</tr>
|
|
27
|
+
{% endfor %}
|
|
28
|
+
</tbody>
|
|
29
|
+
</table>
|
|
30
|
+
</div>
|
|
31
|
+
<p class="mt-1 text-xs text-slate-500">Hover a prompt's name for its text · read after the proposer answered and shown to neither agent · reported, not gated</p>
|
|
32
|
+
</section>
|
|
33
|
+
{% endif %}
|
|
@@ -0,0 +1,24 @@
|
|
|
1
|
+
{# Prior art for _experiment.html. Expects `sub`. #}
|
|
2
|
+
{% if sub.prior_art %}
|
|
3
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
4
|
+
<h3 class="mb-1 text-base font-semibold">Prior art</h3>
|
|
5
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">Published explanations of this unit, or of latents whose weights align with it, looked up after the proposer answered. The proposer never saw them; the reviewer did, as a source of probes. They are other people's claims, not measurements.</p>
|
|
6
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
7
|
+
<table class="w-full text-left text-xs">
|
|
8
|
+
<thead class="bg-slate-50 uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
9
|
+
<tr><th class="px-3 py-2">Latent</th><th class="px-3 py-2">Relation</th><th class="px-3 py-2">Explanation</th><th class="px-3 py-2">By</th></tr>
|
|
10
|
+
</thead>
|
|
11
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
12
|
+
{% for p in sub.prior_art %}
|
|
13
|
+
<tr>
|
|
14
|
+
<td class="px-3 py-1 font-mono"><a class="underline-offset-2 hover:underline" href="{{ p.url }}" rel="noopener">{{ p.dictionary }} F{{ p.feature_idx }}</a></td>
|
|
15
|
+
<td class="px-3 py-1">{% if p.relation == "self" %}this latent{% else %}{{ p.relation }} along the pole ({{ "%+.2f" | format(p.cosine) }}){% endif %}</td>
|
|
16
|
+
<td class="px-3 py-1">{{ p.description }}</td>
|
|
17
|
+
<td class="px-3 py-1 text-slate-500">{{ p.source | capitalize }}, {{ p.explainer }}</td>
|
|
18
|
+
</tr>
|
|
19
|
+
{% endfor %}
|
|
20
|
+
</tbody>
|
|
21
|
+
</table>
|
|
22
|
+
</div>
|
|
23
|
+
</section>
|
|
24
|
+
{% endif %}
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
{# The WordNet specificity check for _experiment.html. Expects `sub` and `what`. #}
|
|
2
|
+
{% if sub.wordnet_specificity %}
|
|
3
|
+
{% set sp = sub.wordnet_specificity %}
|
|
4
|
+
{% macro words(label, items, hint) %}
|
|
5
|
+
<div class="bg-white p-3 dark:bg-slate-950">
|
|
6
|
+
<div class="text-[11px] font-medium uppercase tracking-wide text-slate-500" title="{{ hint }}">{{ label }}</div>
|
|
7
|
+
<p class="mt-1 text-xs leading-relaxed">
|
|
8
|
+
{% for w in items %}<span class="mr-2 whitespace-nowrap"><span class="font-mono">{{ w.word }}</span> <span class="text-slate-500">{{ (w.strength / sp.scale) | num(2) }}</span></span>{% endfor %}
|
|
9
|
+
</p>
|
|
10
|
+
</div>
|
|
11
|
+
{% endmacro %}
|
|
12
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
13
|
+
<h3 class="mb-1 text-base font-semibold">WordNet specificity</h3>
|
|
14
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">This {{ what }} measured on words that WordNet puts under <span class="font-mono">{{ sp.synset }}</span> ({{ sp.lemma }}{{ ", named by the proposer" if sp.source == "proposer" else ", the largest category of the peak words" }}) and on words of its sibling categories, each in {{ sp.carriers | length }} neutral sentences. Peak words are left out. AUROC <strong>{{ sp.auroc | num(2) }}</strong>: the chance that a category word is stronger than a sibling word (0.5: no difference). Active: {{ sp.category_active | pct }} of category words, {{ sp.siblings_active | pct }} of sibling words. Strengths are fractions of the top activation in the scan.</p>
|
|
15
|
+
<div class="grid gap-px overflow-hidden rounded-lg border border-slate-200 bg-slate-200 sm:grid-cols-2 dark:border-slate-800 dark:bg-slate-800">
|
|
16
|
+
{{ words(sp.lemma, sp.category, "Hyponyms of " ~ sp.synset) }}
|
|
17
|
+
{{ words("Siblings", sp.siblings, sp.sibling_synsets | join(", ")) }}
|
|
18
|
+
</div>
|
|
19
|
+
<p class="mt-1 text-xs text-slate-500">Measured after the proposer answered and shown to neither agent · reported, not gated</p>
|
|
20
|
+
</section>
|
|
21
|
+
{% endif %}
|
|
@@ -0,0 +1,34 @@
|
|
|
1
|
+
{# Steering samples for _experiment.html. Expects `sub`, `positive` and `what`. #}
|
|
2
|
+
{% if sub.steering and sub.steering.samples %}
|
|
3
|
+
{% set st = sub.steering %}
|
|
4
|
+
{% set clamp_class = "text-orange-700 dark:text-orange-400" if positive else "text-sky-700 dark:text-sky-400" %}
|
|
5
|
+
{% macro panel(label, text, hint, label_class, baseline=none) %}
|
|
6
|
+
<div class="bg-white p-3 dark:bg-slate-950">
|
|
7
|
+
<div class="text-[11px] font-medium uppercase tracking-wide {{ label_class }}" title="{{ hint }}">{{ label }}</div>
|
|
8
|
+
{% if baseline is not none and text == baseline %}
|
|
9
|
+
<p class="tok mt-1 font-mono text-xs leading-relaxed text-slate-400 dark:text-slate-600" title="Same as the baseline">{{ text }}</p>
|
|
10
|
+
{% else %}
|
|
11
|
+
<p class="tok mt-1 font-mono text-xs leading-relaxed">{{ text }}</p>
|
|
12
|
+
{% endif %}
|
|
13
|
+
</div>
|
|
14
|
+
{% endmacro %}
|
|
15
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
16
|
+
<h3 class="mb-1 text-base font-semibold" title="Turner et al., 2023; Goodfire, Neural Networks Are Already Clean, 2026">Steering</h3>
|
|
17
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">What the model writes after neutral prompts with this {{ what }} ablated, untouched, and clamped at multiples of its top activation in the scan ({{ st.clamp_unit | num(2) }}), at every token, decoding greedily. Drift that appears at ×1–×2 says what the {{ what }} makes the model write; drift only at ×4 is off-distribution. Text no different from the baseline is faded.</p>
|
|
18
|
+
<div class="space-y-4">
|
|
19
|
+
{% for s in st.samples %}
|
|
20
|
+
<figure>
|
|
21
|
+
<figcaption class="mb-1 font-mono text-xs text-slate-500">{{ s.prompt }} …</figcaption>
|
|
22
|
+
<div class="grid gap-px overflow-hidden rounded-lg border border-slate-200 bg-slate-200 sm:grid-cols-2 lg:grid-cols-{{ 2 + st.multiples | length }} dark:border-slate-800 dark:bg-slate-800">
|
|
23
|
+
{{ panel("Ablated", s.ablated, "The " ~ what ~ " ablated at every token", "text-slate-500", s.baseline) }}
|
|
24
|
+
{{ panel("Baseline", s.baseline, "No intervention", "text-slate-500") }}
|
|
25
|
+
{% for m in st.multiples %}
|
|
26
|
+
{{ panel("Clamped ×" ~ ("%g" | format(m)), s.clamped[loop.index0], "Clamped to " ~ ("%.2f" | format(m * st.clamp_unit)) ~ " at every token", clamp_class, s.baseline) }}
|
|
27
|
+
{% endfor %}
|
|
28
|
+
</div>
|
|
29
|
+
</figure>
|
|
30
|
+
{% endfor %}
|
|
31
|
+
</div>
|
|
32
|
+
<p class="mt-1 text-xs text-slate-500">{{ st.max_new_tokens }} new tokens at most · taken after the proposer answered and shown to neither agent · reported, not gated</p>
|
|
33
|
+
</section>
|
|
34
|
+
{% endif %}
|
|
@@ -0,0 +1,62 @@
|
|
|
1
|
+
{# A superposition's members, and the superpositions that a neuron is a member of (builder/superpositions.py). Included from neuron.html. #}
|
|
2
|
+
{% from "_macros.html" import status_badge %}
|
|
3
|
+
{% if neuron.is_superposition and neuron.members %}
|
|
4
|
+
<section class="mt-6" aria-labelledby="members">
|
|
5
|
+
<h2 id="members" class="text-base font-semibold">Members</h2>
|
|
6
|
+
<p class="mt-1 max-w-3xl text-xs text-slate-500">
|
|
7
|
+
This superposition is a sparse direction over the neurons of layer {{ neuron.layer }}. Its strength at a token is the positive part of the projection on that direction.
|
|
8
|
+
Each weight is relative to the largest weight magnitude, and the list is sorted by magnitude.
|
|
9
|
+
{% if neuron.profile %}Method: <code>{{ neuron.profile.method }}</code>. Source: <code>{{ neuron.profile.source }}</code>.{% endif %}
|
|
10
|
+
</p>
|
|
11
|
+
<div class="mt-3 overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
12
|
+
<table class="w-full text-left text-sm">
|
|
13
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
14
|
+
<tr>
|
|
15
|
+
<th class="px-3 py-2">Neuron</th>
|
|
16
|
+
<th class="px-3 py-2 text-right">Relative weight</th>
|
|
17
|
+
</tr>
|
|
18
|
+
</thead>
|
|
19
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
20
|
+
{% for m in neuron.members %}
|
|
21
|
+
<tr>
|
|
22
|
+
<td class="px-3 py-2 font-mono"><a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ root }}{{ m.href }}">N{{ m.neuron_idx }}</a></td>
|
|
23
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums" title="weight {{ m.weight }}">{{ '%+.2f' | format(m.relative) }}</td>
|
|
24
|
+
</tr>
|
|
25
|
+
{% endfor %}
|
|
26
|
+
</tbody>
|
|
27
|
+
</table>
|
|
28
|
+
</div>
|
|
29
|
+
</section>
|
|
30
|
+
{% endif %}
|
|
31
|
+
{% if neuron.part_of %}
|
|
32
|
+
<section class="mt-6" aria-labelledby="superpositions">
|
|
33
|
+
<h2 id="superpositions" class="text-base font-semibold">Superpositions</h2>
|
|
34
|
+
<p class="mt-1 max-w-3xl text-xs text-slate-500">
|
|
35
|
+
{% if not neuron.experiments %}This neuron has no experiment of its own. It is studied only as a member of the superpositions below.{% else %}The studied superpositions that this neuron is a member of.{% endif %}
|
|
36
|
+
The weight is this neuron's weight, relative to the largest weight magnitude in the superposition.
|
|
37
|
+
</p>
|
|
38
|
+
<div class="mt-3 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">Superposition</th>
|
|
43
|
+
<th class="px-3 py-2">Status</th>
|
|
44
|
+
<th class="px-3 py-2 text-right">Weight</th>
|
|
45
|
+
<th class="px-3 py-2">Explanation</th>
|
|
46
|
+
</tr>
|
|
47
|
+
</thead>
|
|
48
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
49
|
+
{% for x in neuron.part_of %}
|
|
50
|
+
{% set sup = x.superposition %}
|
|
51
|
+
<tr class="align-top">
|
|
52
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono"><a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ root }}{{ sup.href }}">L{{ sup.layer }}/{{ sup.label }}</a></td>
|
|
53
|
+
<td class="px-3 py-2">{% if x.standing %}{{ status_badge(x.standing.status) }}{% endif %}</td>
|
|
54
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{{ '%+.2f' | format(x.relative) }}</td>
|
|
55
|
+
<td class="px-3 py-2 text-xs">{% if x.standing %}{{ x.standing.lead.record.submission.natural_language_explanation | truncate(160) }}{% endif %}</td>
|
|
56
|
+
</tr>
|
|
57
|
+
{% endfor %}
|
|
58
|
+
</tbody>
|
|
59
|
+
</table>
|
|
60
|
+
</div>
|
|
61
|
+
</section>
|
|
62
|
+
{% endif %}
|
|
@@ -0,0 +1,88 @@
|
|
|
1
|
+
{# Vocabulary projection for _experiment.html. Expects `v` and `sub`. #}
|
|
2
|
+
{% if sub.vocabulary_lens %}
|
|
3
|
+
{% set vl = sub.vocabulary_lens %}{% set vc = v.vocab %}
|
|
4
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
5
|
+
<h3 class="mb-1 text-base font-semibold">Vocabulary projection</h3>
|
|
6
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">The tokens whose logits this pole's output direction raises and lowers most, read through the final norm and the unembedding.</p>
|
|
7
|
+
{% if vc %}
|
|
8
|
+
<figure>
|
|
9
|
+
<div class="grid gap-x-8 gap-y-6 lg:grid-cols-2">
|
|
10
|
+
{% for panel in vc.panels %}
|
|
11
|
+
<div>
|
|
12
|
+
<h4 class="mb-1 text-sm font-medium" title="{{ panel.hint }}">{{ panel.title }}</h4>
|
|
13
|
+
<svg viewBox="0 0 {{ vc.width }} {{ vc.height }}" class="w-full max-w-[480px]" role="img" aria-label="{{ panel.title }}: z-scores of the {{ panel.rows | selectattr('promoted') | list | length }} most promoted and {{ panel.rows | rejectattr('promoted') | list | length }} most suppressed tokens. Values are in the table below.">
|
|
14
|
+
{% for tx, label in vc.ticks %}
|
|
15
|
+
<line x1="{{ tx }}" y1="0" x2="{{ tx }}" y2="{{ vc.axis_y }}" class="stroke-slate-200 dark:stroke-slate-800" stroke-width="1"/>
|
|
16
|
+
{% endfor %}
|
|
17
|
+
{% for r in panel.rows %}
|
|
18
|
+
<g>
|
|
19
|
+
<title>{{ r.label }}</title>
|
|
20
|
+
<rect x="0" y="{{ r.y - (vc.pitch - vc.bar) / 2 }}" width="{{ vc.width }}" height="{{ vc.pitch }}" fill="transparent"/>
|
|
21
|
+
<path d="{{ r.d }}" class="{{ 'fill-violet-500 dark:fill-violet-400' if r.promoted else 'fill-slate-400 dark:fill-slate-500' }}"/>
|
|
22
|
+
<text x="{{ r.text_x }}" y="{{ r.y + vc.bar - 2 }}" text-anchor="{{ r.anchor }}" class="fill-slate-700 font-mono dark:fill-slate-300" font-size="11" style="white-space: pre">{{ r.text }}</text>
|
|
23
|
+
</g>
|
|
24
|
+
{% endfor %}
|
|
25
|
+
<line x1="{{ vc.zero_x }}" y1="0" x2="{{ vc.zero_x }}" y2="{{ vc.axis_y }}" class="stroke-slate-400" stroke-width="1"/>
|
|
26
|
+
<line x1="0" y1="{{ vc.axis_y + 0.5 }}" x2="{{ vc.width }}" y2="{{ vc.axis_y + 0.5 }}" class="stroke-slate-400" stroke-width="1"/>
|
|
27
|
+
{% for tx, label in vc.ticks %}
|
|
28
|
+
<text x="{{ tx }}" y="{{ vc.axis_y + 14 }}" text-anchor="{{ 'start' if loop.first else 'end' if loop.last else 'middle' }}" class="fill-slate-500 font-mono" font-size="10">{{ label }}</text>
|
|
29
|
+
{% endfor %}
|
|
30
|
+
</svg>
|
|
31
|
+
</div>
|
|
32
|
+
{% endfor %}
|
|
33
|
+
</div>
|
|
34
|
+
<figcaption class="mt-2 flex flex-wrap items-center gap-x-4 gap-y-1 text-xs text-slate-500">
|
|
35
|
+
<span><svg class="inline align-middle" width="12" height="12" aria-hidden="true"><rect x="1" y="2" width="10" height="8" rx="2" class="fill-violet-500 dark:fill-violet-400"/></svg> promotes</span>
|
|
36
|
+
<span><svg class="inline align-middle" width="12" height="12" aria-hidden="true"><rect x="1" y="2" width="10" height="8" rx="2" class="fill-slate-400 dark:fill-slate-500"/></svg> suppresses</span>
|
|
37
|
+
<span>Bar length is the token's z-score over the whole vocabulary, so both lenses share one axis; ␣ marks a space. Hover a bar for its value.</span>
|
|
38
|
+
</figcaption>
|
|
39
|
+
</figure>
|
|
40
|
+
<details class="mt-3">
|
|
41
|
+
<summary class="cursor-pointer text-xs text-slate-500">Values</summary>
|
|
42
|
+
<div class="mt-2 overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
43
|
+
<table class="w-full text-left text-xs">
|
|
44
|
+
<thead class="bg-slate-50 uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
45
|
+
<tr><th class="px-3 py-2">#</th>{% for panel in vc.panels %}<th class="px-3 py-2" colspan="2">{{ panel.title }} promotes</th><th class="px-3 py-2" colspan="2">suppresses</th>{% endfor %}</tr>
|
|
46
|
+
</thead>
|
|
47
|
+
<tbody class="divide-y divide-slate-200 font-mono dark:divide-slate-800">
|
|
48
|
+
{% for row in vc.table %}
|
|
49
|
+
<tr>
|
|
50
|
+
<td class="px-3 py-1 text-slate-500">{{ row.rank }}</td>
|
|
51
|
+
{% for panel in vc.panels %}{% for side in ("promoted", "suppressed") %}{% set cell = row[panel.key ~ "_" ~ side] %}
|
|
52
|
+
<td class="px-3 py-1">{% if cell %}<code class="tok">{{ cell[0] }}</code>{% endif %}</td>
|
|
53
|
+
<td class="px-3 py-1 text-right tabular-nums">{% if cell %}{{ "%+.2f" | format(cell[1]) }}{% endif %}</td>
|
|
54
|
+
{% endfor %}{% endfor %}
|
|
55
|
+
</tr>
|
|
56
|
+
{% endfor %}
|
|
57
|
+
</tbody>
|
|
58
|
+
</table>
|
|
59
|
+
</div>
|
|
60
|
+
</details>
|
|
61
|
+
{% else %}
|
|
62
|
+
<table class="w-full max-w-3xl text-xs">
|
|
63
|
+
<thead class="text-slate-500">
|
|
64
|
+
<tr>
|
|
65
|
+
<th class="pb-1 text-left font-normal"></th>
|
|
66
|
+
<th class="pb-1 text-left font-normal" title="Output direction transported to the final layer with the context-averaged Jacobian (Gurnee et al., 2026)">Jacobian lens</th>
|
|
67
|
+
<th class="pb-1 text-left font-normal" title="Output direction read directly through the unembedding (J = I)">Logit lens</th>
|
|
68
|
+
</tr>
|
|
69
|
+
</thead>
|
|
70
|
+
<tbody class="align-top">
|
|
71
|
+
{% for label, jac, logit in [("Promotes", vl.jacobian_promoted, vl.logit_promoted), ("Suppresses", vl.jacobian_suppressed, vl.logit_suppressed)] %}
|
|
72
|
+
<tr>
|
|
73
|
+
<td class="pr-2 text-slate-500">{{ label }}</td>
|
|
74
|
+
<td class="pr-2">{% for t in jac[:8] %}<code class="tok mr-1 rounded bg-slate-100 px-1 dark:bg-slate-800">{{ t }}</code>{% endfor %}</td>
|
|
75
|
+
<td class="text-slate-500">{% for t in logit[:8] %}<code class="tok mr-1 rounded px-1">{{ t }}</code>{% endfor %}</td>
|
|
76
|
+
</tr>
|
|
77
|
+
{% endfor %}
|
|
78
|
+
</tbody>
|
|
79
|
+
</table>
|
|
80
|
+
<p class="mt-1 text-xs text-slate-500">This record predates vocabulary z-scores, so the tokens are listed without bars.</p>
|
|
81
|
+
{% endif %}
|
|
82
|
+
<p class="mt-2 text-xs text-slate-500">
|
|
83
|
+
Top-10 overlap {{ vl.overlap | pct }} · cos(w, J·w) {{ vl.direction_cosine | num(2) }} ·
|
|
84
|
+
kurtosis {{ vl.jacobian_kurtosis | num(1) }} (Jacobian) / {{ vl.logit_kurtosis | num(1) }} (logit); ≈3 means no specific token set ·
|
|
85
|
+
{{ vl.prompts }} prompts
|
|
86
|
+
</p>
|
|
87
|
+
</section>
|
|
88
|
+
{% endif %}
|
|
@@ -0,0 +1,44 @@
|
|
|
1
|
+
{# A unit's wiring for _experiment.html. Expects `v.wiring` from circuits.wiring_view, `what` and `root`. #}
|
|
2
|
+
{% if v.wiring %}
|
|
3
|
+
<section class="mb-10" data-pagefind-ignore>
|
|
4
|
+
<h3 class="mb-1 text-base font-semibold" title="Elhage et al., 2021; Syed et al., 2023">Wiring</h3>
|
|
5
|
+
<p class="mb-4 max-w-3xl text-sm text-slate-600 dark:text-slate-400">The units that write most directly into this {{ what }}'s input (upstream) and read most directly from its output (downstream), along the residual stream without passing through other units. Linked units have pages in the registry. Reported, not gated.{% if v.wiring_backfilled %} Backfilled after the review by <code>open-interpretability wiring-backfill</code>, from the record's model revision and the review's prompts; it is not part of the record.{% endif %}{% if v.circuit %} <a class="text-emerald-700 underline dark:text-emerald-400" href="{{ root }}{{ v.circuit }}">See its circuit</a>.{% endif %}</p>
|
|
6
|
+
{% for view in v.wiring %}
|
|
7
|
+
{% set w = view.w %}
|
|
8
|
+
<h4 class="mb-1 text-sm font-medium">{{ "From the weights" if w.method == "weights" else "From activations" }}</h4>
|
|
9
|
+
<p class="mb-3 max-w-3xl text-xs text-slate-500">
|
|
10
|
+
{% if w.method == "weights" %}
|
|
11
|
+
Virtual weights with each norm's gain folded in; they rank units but have no absolute scale. Each number is how much one unit reads of the other's write: signed when one vector reads another (+ means a positive activation raises the read), otherwise a norm. Heads count through their OV circuit upstream and their Q, K and V reads downstream; channels are on hover.
|
|
12
|
+
{% else %}
|
|
13
|
+
Averaged over this {{ what }}'s peak on {{ w.prompts }} of the review's prompts where it fired. Upstream: the first-order change in its strength from removing a unit's direct write (+ excites). Downstream: the change in a later neuron's activation, or how far a later head's write moves, when this {{ what }}'s write is removed from that unit's input (one of Q, K or V for a head) alone.
|
|
14
|
+
{% endif %}
|
|
15
|
+
</p>
|
|
16
|
+
<div class="mb-6 grid gap-6 sm:grid-cols-2 lg:grid-cols-4">
|
|
17
|
+
{% for title, rows in view.groups %}
|
|
18
|
+
<div>
|
|
19
|
+
<h5 class="mb-1 text-xs font-medium uppercase tracking-wide text-slate-500">{{ title }}</h5>
|
|
20
|
+
{% if rows is none %}
|
|
21
|
+
<p class="text-xs text-slate-500">Not measured.</p>
|
|
22
|
+
{% elif not rows %}
|
|
23
|
+
<p class="text-xs text-slate-500">None: {{ "no earlier layer" if title.startswith("Upstream") else "no later layer" }}.</p>
|
|
24
|
+
{% else %}
|
|
25
|
+
<table class="w-full font-mono text-xs">
|
|
26
|
+
<tbody>
|
|
27
|
+
{% for row in rows %}
|
|
28
|
+
{% set e = row.e %}
|
|
29
|
+
{% set label = "L" ~ e.layer ~ ("/N" if e.kind == "mlp_neuron" else "/H") ~ e.index %}
|
|
30
|
+
<tr title="{% for c, x in e.channels.items() %}{{ c }} {{ '%+.3g' | format(x) }}{{ ' · ' if not loop.last }}{% endfor %}">
|
|
31
|
+
<td>{% if row.href %}<a class="text-emerald-700 underline dark:text-emerald-400" href="{{ root }}{{ row.href }}">{{ label }}</a>{% else %}{{ label }}{% endif %}</td>
|
|
32
|
+
<td class="text-slate-500">{{ row.via or "" }}</td>
|
|
33
|
+
<td class="text-right tabular-nums">{{ '%+.3g' | format(e.weight) }}</td>
|
|
34
|
+
</tr>
|
|
35
|
+
{% endfor %}
|
|
36
|
+
</tbody>
|
|
37
|
+
</table>
|
|
38
|
+
{% endif %}
|
|
39
|
+
</div>
|
|
40
|
+
{% endfor %}
|
|
41
|
+
</div>
|
|
42
|
+
{% endfor %}
|
|
43
|
+
</section>
|
|
44
|
+
{% endif %}
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
{% extends "base.html" %}
|
|
2
|
+
|
|
3
|
+
{% block title %}About · {{ site_title }}{% endblock %}
|
|
4
|
+
|
|
5
|
+
{% set protocol_url = repo_url ~ "/blob/main/docs/PROTOCOL.md" %}
|
|
6
|
+
{% set link = "text-emerald-700 underline underline-offset-2 dark:text-emerald-400" %}
|
|
7
|
+
|
|
8
|
+
{% block content %}
|
|
9
|
+
<nav class="mb-4 text-xs text-slate-500"><a class="underline" href="{{ root }}index.html">Registry</a> / About</nav>
|
|
10
|
+
<h1 class="text-2xl font-semibold">About</h1>
|
|
11
|
+
|
|
12
|
+
<div class="mt-6 max-w-3xl space-y-8 text-sm leading-relaxed text-slate-700 dark:text-slate-300">
|
|
13
|
+
<section>
|
|
14
|
+
<h2 class="text-lg font-semibold text-slate-900 dark:text-slate-100">Why</h2>
|
|
15
|
+
<p class="mt-2">
|
|
16
|
+
{{ site_title }} is a shared map of what each neuron in an open-weight model does, built in the spirit of
|
|
17
|
+
SETI@home: volunteers give agent time and compute, and every result is checked against the model's weights
|
|
18
|
+
before it joins the map. Each entry is one pole (the positive or negative half) of one native MLP neuron,
|
|
19
|
+
not a latent of a learned dictionary.
|
|
20
|
+
</p>
|
|
21
|
+
<p class="mt-2">
|
|
22
|
+
The project tests a working hypothesis, not an established finding. Goodfire's essay
|
|
23
|
+
<a class="{{ link }}" href="https://already-clean.goodfire.com/"><em>Neural Networks Are Already Clean</em></a>
|
|
24
|
+
(Cammarata, 2026) argues that most apparent polysemanticity in native neurons comes from fixable tooling
|
|
25
|
+
problems rather than superposition. The registry tests that claim one pole at a time. A pole that resists a
|
|
26
|
+
clean explanation is a result too, and every experiment is published whatever its verdict.
|
|
27
|
+
</p>
|
|
28
|
+
</section>
|
|
29
|
+
|
|
30
|
+
<section>
|
|
31
|
+
<h2 class="text-lg font-semibold text-slate-900 dark:text-slate-100">How a verdict is decided</h2>
|
|
32
|
+
<p class="mt-2">
|
|
33
|
+
Agents only write text. Every number in a verdict comes from running the model, so anyone with the weights
|
|
34
|
+
can reproduce it.
|
|
35
|
+
</p>
|
|
36
|
+
<ol class="mt-2 list-decimal space-y-1 pl-5">
|
|
37
|
+
<li><strong>Propose.</strong> A proposer agent scans a corpus for the pole's top-activating text, then writes a
|
|
38
|
+
hypothesis, says whether the pole reacts to tokens it has read or prepares an upcoming one, and predicts how
|
|
39
|
+
much minimal-pair counterfactuals change its activation.</li>
|
|
40
|
+
<li><strong>Review.</strong> A reviewer agent, ideally a different model, writes adversarial probes from the
|
|
41
|
+
hypothesis alone: hard negatives the hypothesis says must not fire, and positive holdouts it says must.
|
|
42
|
+
It never sees a probe's activation before its probe set is final.</li>
|
|
43
|
+
<li><strong>Measure.</strong> The harness runs every probe and counterfactual on the model and applies a fixed
|
|
44
|
+
decision rule. No LLM judgement goes into the verdict.</li>
|
|
45
|
+
</ol>
|
|
46
|
+
<p class="mt-2">
|
|
47
|
+
A verdict is <strong>VERIFIED</strong> only if every gate passes. With near-duplicate probes counted once, there
|
|
48
|
+
must be at least {{ policy.min_hard_negatives }} hard negatives and {{ policy.min_holdouts }} holdouts;
|
|
49
|
+
specificity must be at least {{ "%.2f" | format(policy.min_specificity) }} and sensitivity at least
|
|
50
|
+
{{ "%.2f" | format(policy.min_sensitivity) }}; at least {{ "%.2f" | format(policy.min_counterfactual_reproduction) }}
|
|
51
|
+
of the predicted counterfactual changes must reproduce; the measured mechanism must match the claim; and the
|
|
52
|
+
hypothesis must explain the probes at least as well as trivial token-identity, position and never-active
|
|
53
|
+
baselines. A well-probed hypothesis with balanced accuracy below {{ "%.2f" | format(policy.reject_below_balanced_accuracy) }}
|
|
54
|
+
is <strong>REJECTED</strong>. Every other outcome is <strong>NEEDS_REVISION</strong>, with its failed gates listed.
|
|
55
|
+
</p>
|
|
56
|
+
<p class="mt-2">
|
|
57
|
+
VERIFIED means the hypothesis survived one adversarial probe set and predicts the pole's activation better
|
|
58
|
+
than the trivial baselines on it. It doesn't mean the explanation is complete, that it covers the pole's whole
|
|
59
|
+
activation range, or that the pole has no other role; a later experiment can revise it. Causal sufficiency and
|
|
60
|
+
necessity are measured and shown on each experiment's page, but not yet gated.
|
|
61
|
+
</p>
|
|
62
|
+
<p class="mt-2">
|
|
63
|
+
The full rule is in the <a class="{{ link }}" href="{{ protocol_url }}">audit protocol</a>, currently
|
|
64
|
+
version <span class="font-mono">{{ policy.protocol_version }}</span>. Each verdict stores the policy it was
|
|
65
|
+
decided under, so any record can be re-checked.
|
|
66
|
+
</p>
|
|
67
|
+
</section>
|
|
68
|
+
|
|
69
|
+
<section>
|
|
70
|
+
<h2 class="text-lg font-semibold text-slate-900 dark:text-slate-100">Contribute</h2>
|
|
71
|
+
<p class="mt-2">
|
|
72
|
+
An experiment takes one neuron, proposes what it does, has a reviewer try to break the claim and posts the
|
|
73
|
+
result as a pull request. An agent in Claude Code can drive the whole loop, or you can run each step yourself.
|
|
74
|
+
The <a class="{{ link }}" href="{{ repo_url }}#readme">README</a> covers setup, choosing a neuron and opening
|
|
75
|
+
the PR. Set <span class="font-mono">OPENINTERP_CONTRIBUTOR</span> to a public handle to be credited on the
|
|
76
|
+
<a class="{{ link }}" href="{{ root }}agents/index.html">Agents</a> page.
|
|
77
|
+
</p>
|
|
78
|
+
<p class="mt-2">
|
|
79
|
+
The full registry is open: download <a class="{{ link }}" href="{{ root }}data/registry.parquet">registry.parquet</a>,
|
|
80
|
+
or browse every signed record in the repository's <span class="font-mono">wiki/</span>.
|
|
81
|
+
</p>
|
|
82
|
+
</section>
|
|
83
|
+
|
|
84
|
+
<section>
|
|
85
|
+
<h2 class="text-lg font-semibold text-slate-900 dark:text-slate-100">Licence</h2>
|
|
86
|
+
<p class="mt-2">
|
|
87
|
+
The code is released under the <a class="{{ link }}" href="{{ repo_url }}/blob/main/LICENSE">{{ license }} licence</a>.
|
|
88
|
+
The data (the experiment records, their pages and this site) is released under
|
|
89
|
+
<a class="{{ link }}" href="{{ data_license_url }}">{{ data_license }}</a>: reuse it, and credit Open Interpretability.
|
|
90
|
+
Corpus excerpts on experiment pages keep their source's terms, shown next to them. Questions:
|
|
91
|
+
<a class="{{ link }}" href="mailto:{{ contact_email }}">{{ contact_email }}</a>.
|
|
92
|
+
</p>
|
|
93
|
+
</section>
|
|
94
|
+
</div>
|
|
95
|
+
{% endblock %}
|
|
@@ -0,0 +1,69 @@
|
|
|
1
|
+
{% extends "base.html" %}
|
|
2
|
+
{% from "_macros.html" import status_badge, self_review_badge, pole_mark %}
|
|
3
|
+
|
|
4
|
+
{% block title %}{{ agent.model }} · Agents · {{ site_title }}{% endblock %}
|
|
5
|
+
|
|
6
|
+
{% block content %}
|
|
7
|
+
<nav class="mb-4 text-xs text-slate-500"><a class="underline" href="{{ root }}index.html">Registry</a> / <a class="underline" href="{{ root }}agents/index.html">Agents</a> / {{ agent.model }}</nav>
|
|
8
|
+
<h1 class="break-all font-mono text-2xl font-semibold">{{ agent.model }}</h1>
|
|
9
|
+
<p class="mt-2 text-sm text-slate-600 dark:text-slate-400">Took part in {{ agent.experiments }} experiment{{ "" if agent.experiments == 1 else "s" }}. See <a class="underline" href="{{ root }}agents/index.html">all agents</a> for how the rate and costs are counted.</p>
|
|
10
|
+
{% if agent.other_spellings %}<p class="mt-1 text-sm text-slate-600 dark:text-slate-400">Recorded as {% for s in agent.other_spellings %}<span class="font-mono text-xs">{{ s }}</span>{{ ", " if not loop.last }}{% endfor %}.</p>{% endif %}
|
|
11
|
+
{% if agent.note %}<p class="mt-2 rounded border border-amber-300 bg-amber-50 px-3 py-2 text-sm text-amber-900 dark:border-amber-800 dark:bg-amber-950 dark:text-amber-200">{{ agent.note }}</p>{% endif %}
|
|
12
|
+
|
|
13
|
+
<section class="mt-6">
|
|
14
|
+
<h2 class="mb-3 text-lg font-semibold">By role</h2>
|
|
15
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
16
|
+
<table class="w-full text-left text-sm">
|
|
17
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
18
|
+
<tr>
|
|
19
|
+
<th class="px-3 py-2">Role</th>
|
|
20
|
+
<th class="px-3 py-2 text-right">Experiments</th>
|
|
21
|
+
<th class="px-3 py-2 text-right" title="VERIFIED / experiments, independent reviews only">VERIFIED rate</th>
|
|
22
|
+
<th class="px-3 py-2 text-right" title="self-reviewed experiments: VERIFIED / experiments">Self-reviewed</th>
|
|
23
|
+
<th class="px-3 py-2 text-right" title="experiments priced / experiments with calls in this role">Priced</th>
|
|
24
|
+
<th class="px-3 py-2 text-right">Mean cost</th>
|
|
25
|
+
<th class="px-3 py-2 text-right" title="total cost of the priced experiments / how many of them were VERIFIED">Per VERIFIED</th>
|
|
26
|
+
<th class="px-3 py-2 text-right" title="total cost of the priced experiments">Total</th>
|
|
27
|
+
</tr>
|
|
28
|
+
</thead>
|
|
29
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
30
|
+
{% for s in agent.role_rows %}
|
|
31
|
+
<tr>
|
|
32
|
+
{% include "_agent_role_cells.html" %}
|
|
33
|
+
<td class="px-3 py-2 text-right font-mono tabular-nums">{% if s.costs and s.costs.priced %}${{ "%.2f" | format(s.costs.usd) }}{% else %}–{% endif %}</td>
|
|
34
|
+
</tr>
|
|
35
|
+
{% endfor %}
|
|
36
|
+
</tbody>
|
|
37
|
+
</table>
|
|
38
|
+
</div>
|
|
39
|
+
</section>
|
|
40
|
+
|
|
41
|
+
<section class="mt-8">
|
|
42
|
+
<h2 class="mb-1 text-lg font-semibold">Most recent experiments</h2>
|
|
43
|
+
<p class="mb-3 text-sm text-slate-600 dark:text-slate-400">The latest {{ agent.recent | length }} experiments it took part in, newest first.</p>
|
|
44
|
+
<div class="overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
45
|
+
<table class="w-full text-left text-sm">
|
|
46
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
47
|
+
<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">Role</th><th class="px-3 py-2">Result</th></tr>
|
|
48
|
+
</thead>
|
|
49
|
+
<tbody class="divide-y divide-slate-200 dark:divide-slate-800">
|
|
50
|
+
{% for r in agent.recent %}
|
|
51
|
+
{% set v = r.e.record.verdict %}
|
|
52
|
+
<tr class="align-top hover:bg-slate-50 dark:hover:bg-slate-900">
|
|
53
|
+
<td class="hidden whitespace-nowrap px-3 py-2 font-mono text-xs text-slate-600 sm:table-cell dark:text-slate-400"><time datetime="{{ v.timestamp.strftime('%Y-%m-%dT%H:%M:%SZ') }}">{{ v.timestamp.strftime('%Y-%m-%d') }}</time></td>
|
|
54
|
+
<td class="whitespace-nowrap px-3 py-2 font-mono text-xs">
|
|
55
|
+
<a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ root }}{{ r.href }}">{{ r.unit }}{% if r.pole %}{{ pole_mark(r.pole) }}{% endif %}</a>
|
|
56
|
+
<div class="text-slate-500">{{ r.model_name }}</div>
|
|
57
|
+
</td>
|
|
58
|
+
<td class="px-3 py-2 text-xs">{{ r.roles | join(", ") }}</td>
|
|
59
|
+
<td class="px-3 py-2">
|
|
60
|
+
<div class="mb-1 flex flex-wrap gap-1">{{ status_badge(v.status) }}{% if r.e.record.self_reviewed %} {{ self_review_badge() }}{% endif %}</div>
|
|
61
|
+
{{ r.e.record.submission.natural_language_explanation | truncate(140) }}
|
|
62
|
+
</td>
|
|
63
|
+
</tr>
|
|
64
|
+
{% endfor %}
|
|
65
|
+
</tbody>
|
|
66
|
+
</table>
|
|
67
|
+
</div>
|
|
68
|
+
</section>
|
|
69
|
+
{% endblock %}
|
|
@@ -0,0 +1,80 @@
|
|
|
1
|
+
{% extends "base.html" %}
|
|
2
|
+
|
|
3
|
+
{% block title %}Agents · {{ 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> / Agents</nav>
|
|
7
|
+
<h1 class="text-2xl font-semibold">Agents</h1>
|
|
8
|
+
<p class="mt-2 max-w-3xl text-sm text-slate-600 dark:text-slate-400">
|
|
9
|
+
Every agent model that proposed a hypothesis, reviewed one or judged an explanation, with one row per role.
|
|
10
|
+
The VERIFIED rate is VERIFIED experiments over experiments, with the sample size beside it. It leaves out
|
|
11
|
+
self-reviewed experiments (the proposer's model also wrote the probes), which are counted in their own column,
|
|
12
|
+
since a model passing its own probes inflates the rate. A judge scores explanations after the verdict, so it has no rate.
|
|
13
|
+
Costs cover experiments where every call of that role was priced. All of this is reported, not gated.
|
|
14
|
+
</p>
|
|
15
|
+
|
|
16
|
+
<div class="mt-6 overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
17
|
+
<table class="w-full text-left text-sm">
|
|
18
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
19
|
+
<tr>
|
|
20
|
+
<th class="px-3 py-2">Model</th>
|
|
21
|
+
<th class="px-3 py-2">Role</th>
|
|
22
|
+
<th class="px-3 py-2 text-right">Experiments</th>
|
|
23
|
+
<th class="px-3 py-2 text-right" title="VERIFIED / experiments, independent reviews only">VERIFIED rate</th>
|
|
24
|
+
<th class="px-3 py-2 text-right" title="self-reviewed experiments: VERIFIED / experiments">Self-reviewed</th>
|
|
25
|
+
<th class="px-3 py-2 text-right" title="experiments priced / experiments with calls in this role">Priced</th>
|
|
26
|
+
<th class="px-3 py-2 text-right">Mean cost</th>
|
|
27
|
+
<th class="px-3 py-2 text-right" title="total cost of the priced experiments / how many of them were VERIFIED">Per VERIFIED</th>
|
|
28
|
+
</tr>
|
|
29
|
+
</thead>
|
|
30
|
+
{% for a in agents %}
|
|
31
|
+
<tbody class="border-t border-slate-200 dark:border-slate-800">
|
|
32
|
+
{% for s in a.role_rows %}
|
|
33
|
+
<tr class="hover:bg-slate-50 dark:hover:bg-slate-900">
|
|
34
|
+
{% if loop.first %}<td class="px-3 py-2 align-top font-mono text-xs" rowspan="{{ a.role_rows | length }}"><a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ root }}{{ a.href }}">{{ a.model }}</a></td>{% endif %}
|
|
35
|
+
{% include "_agent_role_cells.html" %}
|
|
36
|
+
</tr>
|
|
37
|
+
{% endfor %}
|
|
38
|
+
</tbody>
|
|
39
|
+
{% else %}
|
|
40
|
+
<tbody><tr><td class="px-3 py-2 text-slate-500" colspan="8">No agents yet.</td></tr></tbody>
|
|
41
|
+
{% endfor %}
|
|
42
|
+
</table>
|
|
43
|
+
</div>
|
|
44
|
+
|
|
45
|
+
<h2 class="mt-10 text-xl font-semibold" id="contributors">Contributors</h2>
|
|
46
|
+
<p class="mt-2 max-w-3xl text-sm text-slate-600 dark:text-slate-400">
|
|
47
|
+
The people who ran the agents, by the public handle they recorded with <code>--contributor</code>
|
|
48
|
+
(or <code>$OPENINTERP_CONTRIBUTOR</code>). A record without one counts for the person who opened the pull request
|
|
49
|
+
that merged it. Each name links to <a class="underline" href="{{ root }}users/index.html">their page</a>.
|
|
50
|
+
Experiments and VERIFIED count the hypotheses they made;
|
|
51
|
+
well-probed rejections are their proposals with strong evidence against them, which are as useful as a VERIFIED result;
|
|
52
|
+
reviews count the experiments they reviewed.
|
|
53
|
+
</p>
|
|
54
|
+
<div class="mt-4 overflow-x-auto rounded-lg border border-slate-200 dark:border-slate-800">
|
|
55
|
+
<table class="w-full text-left text-sm">
|
|
56
|
+
<thead class="bg-slate-50 text-xs uppercase tracking-wide text-slate-500 dark:bg-slate-900">
|
|
57
|
+
<tr>
|
|
58
|
+
<th class="px-3 py-2">Contributor</th>
|
|
59
|
+
<th class="px-3 py-2 text-right" title="hypotheses proposed">Experiments</th>
|
|
60
|
+
<th class="px-3 py-2 text-right">VERIFIED</th>
|
|
61
|
+
<th class="px-3 py-2 text-right" title="proposals with strong evidence against them">Well-probed rejections</th>
|
|
62
|
+
<th class="px-3 py-2 text-right" title="experiments reviewed">Reviews</th>
|
|
63
|
+
</tr>
|
|
64
|
+
</thead>
|
|
65
|
+
<tbody>
|
|
66
|
+
{% for c in contributors %}
|
|
67
|
+
<tr class="border-t border-slate-200 hover:bg-slate-50 dark:border-slate-800 dark:hover:bg-slate-900">
|
|
68
|
+
<td class="px-3 py-2 font-mono text-xs"><a class="text-emerald-700 underline-offset-2 hover:underline dark:text-emerald-400" href="{{ root }}{{ c.href }}">{{ c.handle }}</a></td>
|
|
69
|
+
<td class="px-3 py-2 text-right tabular-nums">{{ c.experiments }}</td>
|
|
70
|
+
<td class="px-3 py-2 text-right tabular-nums">{{ c.verified }}</td>
|
|
71
|
+
<td class="px-3 py-2 text-right tabular-nums">{{ c.conclusive_nulls }}</td>
|
|
72
|
+
<td class="px-3 py-2 text-right tabular-nums">{{ c.reviews }}</td>
|
|
73
|
+
</tr>
|
|
74
|
+
{% else %}
|
|
75
|
+
<tr><td class="px-3 py-2 text-slate-500" colspan="5">No contributors yet.</td></tr>
|
|
76
|
+
{% endfor %}
|
|
77
|
+
</tbody>
|
|
78
|
+
</table>
|
|
79
|
+
</div>
|
|
80
|
+
{% endblock %}
|