Commit ·
3b151ef
0
Parent(s):
Duplicate from AnonymousInterpScience/pythia-jlens-artifacts
Browse filesCo-authored-by: Anonymous <AnonymousInterpScience@users.noreply.huggingface.co>
This view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +35 -0
- ATTRIBUTION.md +48 -0
- LICENSE +25 -0
- README.md +60 -0
- corpora/dataset_revisions.json +111 -0
- corpora/manifest.json +98 -0
- corpora/manifest_ood.json +218 -0
- results/ctrlbf1_control_backfill.json +305 -0
- results/cv1_answer_competence.json +288 -0
- results/cv2_position_support.json +299 -0
- results/cv3_margins_410m.json +473 -0
- results/cv4_phase1_capability.json +719 -0
- results/cv5_metric_sensitivity_410m.json +296 -0
- results/cv6/cv6_1.4b_n25.json +1336 -0
- results/cv6/cv6_2.8b_n25.json +1416 -0
- results/cv6/cv6_c0_scorer_equivalence.json +171 -0
- results/cv6/ht_cache_1.4b.pt +3 -0
- results/cv6/ht_cache_2.8b.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Github_1.4b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Github_1.4b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Github_1.4b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Github_2.8b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Github_2.8b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Github_2.8b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Pile-CC_1.4b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Pile-CC_1.4b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Pile-CC_1.4b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Pile-CC_2.8b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Pile-CC_2.8b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Pile-CC_2.8b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_StackExchange_1.4b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_StackExchange_1.4b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_StackExchange_1.4b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_StackExchange_2.8b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_StackExchange_2.8b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_StackExchange_2.8b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_1.4b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_1.4b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_1.4b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_2.8b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_2.8b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_2.8b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Wikipedia_en_1.4b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Wikipedia_en_1.4b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Wikipedia_en_1.4b_n25_s2.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Wikipedia_en_2.8b_n25_s0.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Wikipedia_en_2.8b_n25_s1.pt +3 -0
- results/cv6/lenses/lens_INSTREAM_Wikipedia_en_2.8b_n25_s2.pt +3 -0
- results/cv6_per_family_ladder.json +1238 -0
- results/cv7/cv7_1b_n200.json +821 -0
.gitattributes
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
| 2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
| 3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
| 7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
| 8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
| 10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
| 11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
| 12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
| 13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
| 14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
| 15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
| 16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
| 17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
| 18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
| 19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
| 20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
| 21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
| 22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
| 23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
| 24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
| 25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
| 27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
| 28 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
| 29 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
| 30 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
| 31 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
| 32 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
| 33 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
ATTRIBUTION.md
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ATTRIBUTION
|
| 2 |
+
|
| 3 |
+
This repository is **CC BY 4.0 without qualification**. It contains only what this work produced.
|
| 4 |
+
|
| 5 |
+
## What is here
|
| 6 |
+
|
| 7 |
+
| what | size | licence |
|
| 8 |
+
|---|---|---|
|
| 9 |
+
| `results/lenses/**` — 382 fitted operators + sidecars | 16.2 GB | CC BY 4.0 |
|
| 10 |
+
| `results/**.json`, `results/**.npz` — experimental records | ~27 MB | CC BY 4.0 |
|
| 11 |
+
| `corpora/manifest.json`, `corpora/manifest_ood.json`, `corpora/dataset_revisions.json` | 12 KB | CC BY 4.0 |
|
| 12 |
+
|
| 13 |
+
Derived from **Pythia** (EleutherAI, Apache-2.0) via the **Jacobian-lens** reference
|
| 14 |
+
implementation (Apache-2.0, pinned `581d398`). Neither restricts these derivatives.
|
| 15 |
+
|
| 16 |
+
## What is deliberately NOT here, and how to get it
|
| 17 |
+
|
| 18 |
+
The eleven corpus plaintext files are **not distributed**. They are third-party text under
|
| 19 |
+
upstream terms; `Wikipedia_en.jsonl` and `OOD_Wikipedia_2023.jsonl` are Wikipedia-derived and
|
| 20 |
+
carry **CC BY-SA**, which is share-alike and cannot be relicensed as CC BY 4.0. Shipping them
|
| 21 |
+
would have forced a mixed-licence repository and made the CC BY 4.0 claim above false.
|
| 22 |
+
|
| 23 |
+
Rebuild the pool from its public sources:
|
| 24 |
+
|
| 25 |
+
```bash
|
| 26 |
+
bash repro/06_build_corpora.sh # builds, then hash-verifies against the manifests
|
| 27 |
+
bash repro/06_build_corpora.sh --verify # verify a pool you already have
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
`corpora/manifest.json` and `corpora/manifest_ood.json` ship here and carry, per corpus: the
|
| 31 |
+
SHA-256 of the exact file used, the document count, the source dataset id and config, and the
|
| 32 |
+
three disjoint seed-block index ranges. A rebuild that matches those hashes is the same pool.
|
| 33 |
+
|
| 34 |
+
**Measured 2026-08-31, rebuilding from scratch on a clean machine:**
|
| 35 |
+
|
| 36 |
+
| corpus | rebuild |
|
| 37 |
+
|---|---|
|
| 38 |
+
| Pile-CC, StackExchange, Wikipedia_en, Github, USPTO_Backgrounds | **5/5 bit-identical** |
|
| 39 |
+
| OOD_CommonPile, TRAP_FineWeb, CONTROL_PubMed_2023 | **3/3 bit-identical** |
|
| 40 |
+
| `corpora/manifest.json` regenerated | **byte-identical** |
|
| 41 |
+
| OOD_Wikipedia_2023, OOD_arXiv_2023 | not re-run (53 MB / 110 MB of streaming); both carry SHA-256 in the manifest and are checked by the script |
|
| 42 |
+
|
| 43 |
+
**One caveat, stated plainly.** `src/build_corpora.py` calls `load_dataset(
|
| 44 |
+
"monology/pile-uncopyrighted", ...)` with **no `revision=` pin**, so an upstream change to that
|
| 45 |
+
dataset would produce a different pool. The result above is evidence the upstream had not moved
|
| 46 |
+
as of that date — not a guarantee it never will. The manifest is the check: a mismatch means the
|
| 47 |
+
upstream moved, which is a finding to report, not a failure to work around. The OOD panel is more
|
| 48 |
+
stable by construction because its configs are date-stamped (`20231101.en`, `2024-06`).
|
LICENSE
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Creative Commons Attribution 4.0 International (CC BY 4.0)
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2026 The Authors
|
| 4 |
+
|
| 5 |
+
Everything in this repository is licensed under CC BY 4.0:
|
| 6 |
+
https://creativecommons.org/licenses/by/4.0/
|
| 7 |
+
|
| 8 |
+
That claim is unqualified, and it is unqualified BY CONSTRUCTION rather than by assertion: this
|
| 9 |
+
repository contains only artifacts this work produced.
|
| 10 |
+
|
| 11 |
+
results/lenses/** fitted Jacobian operators (.pt) and their provenance sidecars
|
| 12 |
+
results/**.json, **.npz every experimental record and derived statistic
|
| 13 |
+
corpora/manifest*.json the build record of the corpus pool: per-file SHA-256, document
|
| 14 |
+
counts, source dataset id and config, disjoint seed-block ranges
|
| 15 |
+
|
| 16 |
+
NO THIRD-PARTY CORPUS TEXT IS DISTRIBUTED HERE. The eleven corpus files this work used are
|
| 17 |
+
third-party text under upstream terms, and two are Wikipedia-derived and carry CC BY-SA, a
|
| 18 |
+
share-alike licence that cannot be relicensed as CC BY 4.0. Rather than qualify this licence,
|
| 19 |
+
the pool is withheld in full and rebuilt from its public sources by the code repository's
|
| 20 |
+
`repro/06_build_corpora.sh`. The manifests above are this work's own record of that pool and
|
| 21 |
+
are what make the rebuild verifiable.
|
| 22 |
+
|
| 23 |
+
The operators are derived from EleutherAI's Pythia suite (Apache-2.0) using the Jacobian-lens
|
| 24 |
+
reference implementation (Apache-2.0, pinned at 581d398). Neither upstream licence restricts the
|
| 25 |
+
licensing of the derived operators recorded here.
|
README.md
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- interpretability
|
| 5 |
+
- jacobian-lens
|
| 6 |
+
- pythia
|
| 7 |
+
- reproducibility
|
| 8 |
+
library_name: pytorch
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Corpus and context dependence of the Jacobian lens — artifact mirror
|
| 12 |
+
|
| 13 |
+
Fitted operators, experimental records and the corpus pool for the accompanying paper.
|
| 14 |
+
|
| 15 |
+
**Licence.** Everything here is **CC BY 4.0**, without qualification — see `LICENSE`. That is
|
| 16 |
+
true by construction, not by assertion: this repository contains only artifacts this work
|
| 17 |
+
produced. **No third-party corpus text is distributed here.** The corpus pool is rebuilt from its
|
| 18 |
+
public sources by `repro/06_build_corpora.sh` and verified against the manifests that ship here;
|
| 19 |
+
8 of 11 corpora rebuilt bit-identically when measured. `ATTRIBUTION.md` has the detail.
|
| 20 |
+
|
| 21 |
+
## Layout
|
| 22 |
+
|
| 23 |
+
```
|
| 24 |
+
results/lenses/e28/ 166 operators + sidecars corpus x N ladder
|
| 25 |
+
results/lenses/plad/ 53 operators + sidecars per-layer activation dumps
|
| 26 |
+
results/lenses/ladder/ 15 operators + sidecars canonical model ladder (70m ... 2.8b)
|
| 27 |
+
results/lenses/misc/ 8 operators smokes and one-offs
|
| 28 |
+
results/<17 experiment dirs> per-experiment operators and records
|
| 29 |
+
results/*.json 190 experimental records
|
| 30 |
+
corpora/ 3 provenance manifests (NO corpus text — rebuild it, see ATTRIBUTION.md)
|
| 31 |
+
```
|
| 32 |
+
|
| 33 |
+
Every `.pt` is a corpus-averaged Jacobian `J_l = E_x[dh_final/dh_l]` fitted with the reference
|
| 34 |
+
implementation at pin `581d398`, fp32, TF32 off, uncentered.
|
| 35 |
+
|
| 36 |
+
## Reproducing
|
| 37 |
+
|
| 38 |
+
Every statistic in the paper recomputes from `results/**.json` alone — no operators, no GPU:
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
git clone <code repository>
|
| 42 |
+
uv sync && bash repro/01_setup_local.sh
|
| 43 |
+
bash repro/04_fetch_results.sh # pulls results/**.json from here, ~7 MB
|
| 44 |
+
bash repro/RUN_ALL.sh --tier T0 # 12 modules
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
`--tier T1` re-scores the stored operators on CPU and needs the `.pt` from this repo
|
| 48 |
+
(`bash repro/03_fetch_artifacts.sh --all`). `--tier T2` refits on GPU.
|
| 49 |
+
|
| 50 |
+
## Provenance
|
| 51 |
+
|
| 52 |
+
Each results file carries a provenance block: the producing script and its SHA-256, the commit,
|
| 53 |
+
the argv, and the SHA-256 of every declared input. `payload_sha256` covers the record **excluding**
|
| 54 |
+
the provenance block, so it is stable across reruns that change only environment metadata — rerun
|
| 55 |
+
and compare it, and a match means no number moved.
|
| 56 |
+
|
| 57 |
+
No hostname is recorded. Input hashes are the values at write time; where a record was later
|
| 58 |
+
redacted, `payload_sha256_before_redaction` retains the original so both remain verifiable.
|
| 59 |
+
|
| 60 |
+
No corpus plaintext is distributed — see `ATTRIBUTION.md` for the rebuild.
|
corpora/dataset_revisions.json
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_README": [
|
| 3 |
+
"Dataset revisions pinned 2026-09-01. THE SOURCE OF TRUTH for every corpus build.",
|
| 4 |
+
"",
|
| 5 |
+
"WHY. src/build_corpora.py called load_dataset() with no revision=, so the pool depended on",
|
| 6 |
+
"whatever the upstream dataset was on the day you ran it. docs/reproducibility/RELEASE.md",
|
| 7 |
+
"finding B0 recorded that and concluded the pool was not regenerable. A rebuild on 2026-08-31",
|
| 8 |
+
"in fact reproduced 8 of 11 corpora BIT-IDENTICALLY -- evidence the upstream had not moved, not",
|
| 9 |
+
"a guarantee it never would. These pins turn that evidence into a guarantee.",
|
| 10 |
+
"",
|
| 11 |
+
"`sha` is the dataset repo commit on the Hub. Pass it as revision= to load_dataset().",
|
| 12 |
+
"`config` is load_dataset()'s SECOND POSITIONAL argument (the config/subset name), NOT a locale;",
|
| 13 |
+
"null means the dataset declares no configs and none may be passed."
|
| 14 |
+
],
|
| 15 |
+
"monology/pile-uncopyrighted": {
|
| 16 |
+
"used_by": "src/build_corpora.py",
|
| 17 |
+
"config": null,
|
| 18 |
+
"split": "train",
|
| 19 |
+
"supplies": [
|
| 20 |
+
"Pile-CC",
|
| 21 |
+
"StackExchange",
|
| 22 |
+
"Wikipedia (en)",
|
| 23 |
+
"Github",
|
| 24 |
+
"USPTO Backgrounds"
|
| 25 |
+
],
|
| 26 |
+
"sha": "3be90335b66f24456a5d6659d9c8d208c0357119",
|
| 27 |
+
"pinned_on": "2026-09-01"
|
| 28 |
+
},
|
| 29 |
+
"wikimedia/wikipedia": {
|
| 30 |
+
"used_by": "src/build_ood_corpora.py",
|
| 31 |
+
"config": "20231101.en",
|
| 32 |
+
"split": "train",
|
| 33 |
+
"supplies": [
|
| 34 |
+
"OOD_Wikipedia_2023"
|
| 35 |
+
],
|
| 36 |
+
"sha": "b04c8d1ceb2f5cd4588862100d08de323dccfbaa",
|
| 37 |
+
"pinned_on": "2026-09-01"
|
| 38 |
+
},
|
| 39 |
+
"neuralwork/arxiver": {
|
| 40 |
+
"used_by": "src/build_ood_corpora.py",
|
| 41 |
+
"config": null,
|
| 42 |
+
"split": "train",
|
| 43 |
+
"supplies": [
|
| 44 |
+
"OOD_arXiv_2023"
|
| 45 |
+
],
|
| 46 |
+
"sha": "698a6662e77fd5dd45dbbec988abc8123e5fa086",
|
| 47 |
+
"pinned_on": "2026-09-01"
|
| 48 |
+
},
|
| 49 |
+
"RealTimeData/bbc_news_alltime": {
|
| 50 |
+
"used_by": "src/build_ood_corpora.py",
|
| 51 |
+
"config": "2024-06",
|
| 52 |
+
"split": "train",
|
| 53 |
+
"supplies": [
|
| 54 |
+
"OOD_News_2024"
|
| 55 |
+
],
|
| 56 |
+
"sha": "8dd1ecdc92ac43f9c04a3da3e945537dbb08179b",
|
| 57 |
+
"pinned_on": "2026-09-01"
|
| 58 |
+
},
|
| 59 |
+
"common-pile/comma_v0.1_training_dataset": {
|
| 60 |
+
"used_by": "src/build_ood_corpora.py",
|
| 61 |
+
"config": null,
|
| 62 |
+
"split": "train",
|
| 63 |
+
"supplies": [
|
| 64 |
+
"OOD_CommonPile"
|
| 65 |
+
],
|
| 66 |
+
"sha": "5afc546db324e7f39f297ba757c9a60547151e7c",
|
| 67 |
+
"pinned_on": "2026-09-01"
|
| 68 |
+
},
|
| 69 |
+
"HuggingFaceFW/fineweb": {
|
| 70 |
+
"used_by": "src/build_ood_corpora.py",
|
| 71 |
+
"config": "sample-10BT",
|
| 72 |
+
"split": "train",
|
| 73 |
+
"supplies": [
|
| 74 |
+
"TRAP_FineWeb"
|
| 75 |
+
],
|
| 76 |
+
"sha": "9bb295ddab0e05d785b879661af7260fed5140fc",
|
| 77 |
+
"pinned_on": "2026-09-01"
|
| 78 |
+
},
|
| 79 |
+
"MedRAG/pubmed": {
|
| 80 |
+
"used_by": "src/build_ood_corpora.py",
|
| 81 |
+
"config": null,
|
| 82 |
+
"split": "train",
|
| 83 |
+
"supplies": [
|
| 84 |
+
"CONTROL_PubMed_2023"
|
| 85 |
+
],
|
| 86 |
+
"sha": "33da3593d5756bc04c8909f170003c0b14197957",
|
| 87 |
+
"pinned_on": "2026-09-01"
|
| 88 |
+
},
|
| 89 |
+
"EleutherAI/pile-deduped-pythia-preshuffled": {
|
| 90 |
+
"used_by": "experiments/t35_containment.py",
|
| 91 |
+
"config": null,
|
| 92 |
+
"split": null,
|
| 93 |
+
"supplies": [
|
| 94 |
+
"the Pythia training stream itself, for E35/M1 containment"
|
| 95 |
+
],
|
| 96 |
+
"sha": "4647773ea142ab1ff5694602fa104bbf49088408",
|
| 97 |
+
"pinned_on": "2026-09-01"
|
| 98 |
+
},
|
| 99 |
+
"_DERIVED_CORPORA": {
|
| 100 |
+
"_note": "Corpora produced from another corpus in this tree, not loaded from the Hub. They have NO dataset revision of their own by construction. Their full derivation rule and parent hash live in corpora/manifest_ood.json. Recorded here so this file's source-of-truth claim covers every corpus on disk.",
|
| 101 |
+
"OOD_News_2024_dedup": {
|
| 102 |
+
"derived_from": "OOD_News_2024",
|
| 103 |
+
"parent_upstream_repo": "RealTimeData/bbc_news_alltime",
|
| 104 |
+
"parent_sha": "8dd1ecdc92ac43f9c04a3da3e945537dbb08179b",
|
| 105 |
+
"rule": "exact-text dedup, first occurrence kept, parent order preserved (2400 -> 1598)",
|
| 106 |
+
"full_record": "corpora/manifest_ood.json -> OOD_News_2024_dedup.derived_from",
|
| 107 |
+
"revision": null,
|
| 108 |
+
"revision_note": "INTENTIONALLY null \u2014 derived, not loaded. Never fabricate one."
|
| 109 |
+
}
|
| 110 |
+
}
|
| 111 |
+
}
|
corpora/manifest.json
ADDED
|
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"Pile-CC": {
|
| 3 |
+
"file": "Pile-CC.jsonl",
|
| 4 |
+
"n": 2400,
|
| 5 |
+
"sha256": "abbb9f19a6eb6f4c12fb382735832ceec271fece15cfebcf0fe838765d4b1e74",
|
| 6 |
+
"seed_blocks": [
|
| 7 |
+
[
|
| 8 |
+
0,
|
| 9 |
+
800
|
| 10 |
+
],
|
| 11 |
+
[
|
| 12 |
+
800,
|
| 13 |
+
1600
|
| 14 |
+
],
|
| 15 |
+
[
|
| 16 |
+
1600,
|
| 17 |
+
2400
|
| 18 |
+
]
|
| 19 |
+
]
|
| 20 |
+
},
|
| 21 |
+
"StackExchange": {
|
| 22 |
+
"file": "StackExchange.jsonl",
|
| 23 |
+
"n": 2400,
|
| 24 |
+
"sha256": "bc25e2d5775e9719ebcbdaa3bd115d827f72ebf11f47fedb377f8b0a69a28971",
|
| 25 |
+
"seed_blocks": [
|
| 26 |
+
[
|
| 27 |
+
0,
|
| 28 |
+
800
|
| 29 |
+
],
|
| 30 |
+
[
|
| 31 |
+
800,
|
| 32 |
+
1600
|
| 33 |
+
],
|
| 34 |
+
[
|
| 35 |
+
1600,
|
| 36 |
+
2400
|
| 37 |
+
]
|
| 38 |
+
]
|
| 39 |
+
},
|
| 40 |
+
"Wikipedia (en)": {
|
| 41 |
+
"file": "Wikipedia_en.jsonl",
|
| 42 |
+
"n": 2400,
|
| 43 |
+
"sha256": "b30f4cef11add560f7df8ed07a613e3fcc6a4dda865975bec15f2952253bda77",
|
| 44 |
+
"seed_blocks": [
|
| 45 |
+
[
|
| 46 |
+
0,
|
| 47 |
+
800
|
| 48 |
+
],
|
| 49 |
+
[
|
| 50 |
+
800,
|
| 51 |
+
1600
|
| 52 |
+
],
|
| 53 |
+
[
|
| 54 |
+
1600,
|
| 55 |
+
2400
|
| 56 |
+
]
|
| 57 |
+
]
|
| 58 |
+
},
|
| 59 |
+
"Github": {
|
| 60 |
+
"file": "Github.jsonl",
|
| 61 |
+
"n": 2400,
|
| 62 |
+
"sha256": "f2bb459d10cad6c13a8b66c86191bc43df4393ba5b722a5315ea878fad2d1a26",
|
| 63 |
+
"seed_blocks": [
|
| 64 |
+
[
|
| 65 |
+
0,
|
| 66 |
+
800
|
| 67 |
+
],
|
| 68 |
+
[
|
| 69 |
+
800,
|
| 70 |
+
1600
|
| 71 |
+
],
|
| 72 |
+
[
|
| 73 |
+
1600,
|
| 74 |
+
2400
|
| 75 |
+
]
|
| 76 |
+
]
|
| 77 |
+
},
|
| 78 |
+
"USPTO Backgrounds": {
|
| 79 |
+
"file": "USPTO_Backgrounds.jsonl",
|
| 80 |
+
"n": 2400,
|
| 81 |
+
"sha256": "b09ac74bdc8b169ee5c9bea6be481262f743947a92107164f30eca2346e3ccb9",
|
| 82 |
+
"seed_blocks": [
|
| 83 |
+
[
|
| 84 |
+
0,
|
| 85 |
+
800
|
| 86 |
+
],
|
| 87 |
+
[
|
| 88 |
+
800,
|
| 89 |
+
1600
|
| 90 |
+
],
|
| 91 |
+
[
|
| 92 |
+
1600,
|
| 93 |
+
2400
|
| 94 |
+
]
|
| 95 |
+
]
|
| 96 |
+
},
|
| 97 |
+
"_note": "seed_blocks are DISJOINT index ranges; seed k draws its N prompts from block k, so seeds share no prompts and the seed variance is a genuine resampling variance rather than an overlap artifact"
|
| 98 |
+
}
|
corpora/manifest_ood.json
ADDED
|
@@ -0,0 +1,218 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"OOD_Wikipedia_2023": {
|
| 3 |
+
"file": "OOD_Wikipedia_2023.jsonl",
|
| 4 |
+
"n": 2400,
|
| 5 |
+
"sha256": "57fb26cf3997b571543281efe3330efcbab1f37e97f99eb6df4deaa027d98550",
|
| 6 |
+
"seed_blocks": [
|
| 7 |
+
[
|
| 8 |
+
0,
|
| 9 |
+
800
|
| 10 |
+
],
|
| 11 |
+
[
|
| 12 |
+
800,
|
| 13 |
+
1600
|
| 14 |
+
],
|
| 15 |
+
[
|
| 16 |
+
1600,
|
| 17 |
+
2400
|
| 18 |
+
]
|
| 19 |
+
],
|
| 20 |
+
"hf": {
|
| 21 |
+
"path": "wikimedia/wikipedia",
|
| 22 |
+
"name": "20231101.en",
|
| 23 |
+
"split": "train"
|
| 24 |
+
},
|
| 25 |
+
"text_field": "text",
|
| 26 |
+
"claim": "TEMPORAL. Same source family as the Pile's Wikipedia component, but a 2023-11 snapshot. Revisions written after the ~2020 Pile snapshot cannot have been in it. Confounds: three years of new entities, events and editorial convention move with the exclusion. This is external review's PRIMARY recommended panel because it holds language and register nearly fixed.",
|
| 27 |
+
"scanned": 2436,
|
| 28 |
+
"status": "CANDIDATE \u2014 outside-Pile is a HYPOTHESIS to be measured by M1, not a property asserted here"
|
| 29 |
+
},
|
| 30 |
+
"OOD_arXiv_2023": {
|
| 31 |
+
"file": "OOD_arXiv_2023.jsonl",
|
| 32 |
+
"n": 2400,
|
| 33 |
+
"sha256": "565f3541ef189ba624ff56373d7d3df0b0341020578b0c6cffc6391b332ff4af",
|
| 34 |
+
"seed_blocks": [
|
| 35 |
+
[
|
| 36 |
+
0,
|
| 37 |
+
800
|
| 38 |
+
],
|
| 39 |
+
[
|
| 40 |
+
800,
|
| 41 |
+
1600
|
| 42 |
+
],
|
| 43 |
+
[
|
| 44 |
+
1600,
|
| 45 |
+
2400
|
| 46 |
+
]
|
| 47 |
+
],
|
| 48 |
+
"hf": {
|
| 49 |
+
"path": "neuralwork/arxiver",
|
| 50 |
+
"split": "train"
|
| 51 |
+
},
|
| 52 |
+
"text_field": "markdown",
|
| 53 |
+
"claim": "TEMPORAL. 2023 arXiv papers; the Pile's arXiv component predates the snapshot. Stable LaTeX/scientific register. Confounds: new terminology, fields, authors; heavy template reuse means short-k overlap will be high regardless of exposure.",
|
| 54 |
+
"scanned": 2401,
|
| 55 |
+
"status": "CANDIDATE \u2014 outside-Pile is a HYPOTHESIS to be measured by M1, not a property asserted here"
|
| 56 |
+
},
|
| 57 |
+
"OOD_News_2024": {
|
| 58 |
+
"file": "OOD_News_2024.jsonl",
|
| 59 |
+
"n": 2400,
|
| 60 |
+
"sha256": "45aaab463d8987725fe9b7edb5bd316a29c2d547a5a1ffd2c54f41df3be1df80",
|
| 61 |
+
"seed_blocks": [
|
| 62 |
+
[
|
| 63 |
+
0,
|
| 64 |
+
800
|
| 65 |
+
],
|
| 66 |
+
[
|
| 67 |
+
800,
|
| 68 |
+
1600
|
| 69 |
+
],
|
| 70 |
+
[
|
| 71 |
+
1600,
|
| 72 |
+
2400
|
| 73 |
+
]
|
| 74 |
+
],
|
| 75 |
+
"hf": {
|
| 76 |
+
"path": "RealTimeData/bbc_news_alltime",
|
| 77 |
+
"name": "2024-06",
|
| 78 |
+
"split": "train"
|
| 79 |
+
},
|
| 80 |
+
"text_field": "content",
|
| 81 |
+
"claim": "TEMPORAL. June 2024 news, four years past the snapshot. Ordinary English, so it tests transport without immediately confounding on script or competence. Confounds: news register, named entities, and possible syndication into web crawls.",
|
| 82 |
+
"scanned": 2930,
|
| 83 |
+
"status": "CANDIDATE \u2014 outside-Pile is a HYPOTHESIS to be measured by M1, not a property asserted here"
|
| 84 |
+
},
|
| 85 |
+
"OOD_CommonPile": {
|
| 86 |
+
"file": "OOD_CommonPile.jsonl",
|
| 87 |
+
"n": 2400,
|
| 88 |
+
"sha256": "0d7ca5e94b4b1891e1d99a310b263806f370bae861decdfc3cf0f1ebb7a755e4",
|
| 89 |
+
"seed_blocks": [
|
| 90 |
+
[
|
| 91 |
+
0,
|
| 92 |
+
800
|
| 93 |
+
],
|
| 94 |
+
[
|
| 95 |
+
800,
|
| 96 |
+
1600
|
| 97 |
+
],
|
| 98 |
+
[
|
| 99 |
+
1600,
|
| 100 |
+
2400
|
| 101 |
+
]
|
| 102 |
+
],
|
| 103 |
+
"hf": {
|
| 104 |
+
"path": "common-pile/comma_v0.1_training_dataset",
|
| 105 |
+
"split": "train"
|
| 106 |
+
},
|
| 107 |
+
"text_field": "text",
|
| 108 |
+
"claim": "PROVENANCE. Assembled later from openly-licensed and public-domain sources, covering domains not deliberately represented in the Pile. Tests whether the effect is about corpus construction/licensing rather than topic. Confounds: licence filtering changes the source population; some documents may still be older web material.",
|
| 109 |
+
"scanned": 3069,
|
| 110 |
+
"status": "CANDIDATE \u2014 outside-Pile is a HYPOTHESIS to be measured by M1, not a property asserted here"
|
| 111 |
+
},
|
| 112 |
+
"TRAP_FineWeb": {
|
| 113 |
+
"file": "TRAP_FineWeb.jsonl",
|
| 114 |
+
"n": 2400,
|
| 115 |
+
"sha256": "6164a3cc1113f6c3b90fd5f67e3eba2e80324a69d9d4c0ae671aefdefc4b9260",
|
| 116 |
+
"seed_blocks": [
|
| 117 |
+
[
|
| 118 |
+
0,
|
| 119 |
+
800
|
| 120 |
+
],
|
| 121 |
+
[
|
| 122 |
+
800,
|
| 123 |
+
1600
|
| 124 |
+
],
|
| 125 |
+
[
|
| 126 |
+
1600,
|
| 127 |
+
2400
|
| 128 |
+
]
|
| 129 |
+
],
|
| 130 |
+
"hf": {
|
| 131 |
+
"path": "HuggingFaceFW/fineweb",
|
| 132 |
+
"name": "sample-10BT",
|
| 133 |
+
"split": "train"
|
| 134 |
+
},
|
| 135 |
+
"text_field": "text",
|
| 136 |
+
"claim": "DELIBERATE POSITIVE CONTROL, NOT A CLEAN Q. FineWeb is Common-Crawl-derived and Pile-CC is too, so it can contain the same pages, later versions of them, or mirrors. Both external review reviews name CC derivatives as a trap. Included so M1 has something it MUST flag as overlapping: if FineWeb does not show high containment, M1 is broken.",
|
| 137 |
+
"scanned": 2708,
|
| 138 |
+
"status": "CANDIDATE \u2014 outside-Pile is a HYPOTHESIS to be measured by M1, not a property asserted here"
|
| 139 |
+
},
|
| 140 |
+
"CONTROL_PubMed_2023": {
|
| 141 |
+
"file": "CONTROL_PubMed_2023.jsonl",
|
| 142 |
+
"n": 2400,
|
| 143 |
+
"sha256": "cffa12d20484b15c2430fc402a994632a9a27595454d02b457b94ec80c27856a",
|
| 144 |
+
"seed_blocks": [
|
| 145 |
+
[
|
| 146 |
+
0,
|
| 147 |
+
800
|
| 148 |
+
],
|
| 149 |
+
[
|
| 150 |
+
800,
|
| 151 |
+
1600
|
| 152 |
+
],
|
| 153 |
+
[
|
| 154 |
+
1600,
|
| 155 |
+
2400
|
| 156 |
+
]
|
| 157 |
+
],
|
| 158 |
+
"hf": {
|
| 159 |
+
"path": "MedRAG/pubmed",
|
| 160 |
+
"split": "train"
|
| 161 |
+
},
|
| 162 |
+
"text_field": "content",
|
| 163 |
+
"claim": "AMBIGUOUS BY DESIGN. The Pile contains PubMed Central, so this is source-INCLUDED; only post-snapshot articles would be temporally excluded and this dump is not date-filtered. Expected to land BETWEEN the Pile components and the temporal panels, which makes it a useful mid-scale reference point for whatever divergence metric we fit.",
|
| 164 |
+
"scanned": 3161,
|
| 165 |
+
"status": "CANDIDATE \u2014 outside-Pile is a HYPOTHESIS to be measured by M1, not a property asserted here"
|
| 166 |
+
},
|
| 167 |
+
"OOD_News_2024_dedup": {
|
| 168 |
+
"file": "OOD_News_2024_dedup.jsonl",
|
| 169 |
+
"n": 1598,
|
| 170 |
+
"sha256": "713cdbcf4512208f4cb3cac831d2a4eeb76ae0fe577d215473fd71919d022529",
|
| 171 |
+
"seed_blocks": [
|
| 172 |
+
[
|
| 173 |
+
0,
|
| 174 |
+
532
|
| 175 |
+
],
|
| 176 |
+
[
|
| 177 |
+
532,
|
| 178 |
+
1064
|
| 179 |
+
],
|
| 180 |
+
[
|
| 181 |
+
1064,
|
| 182 |
+
1596
|
| 183 |
+
]
|
| 184 |
+
],
|
| 185 |
+
"seed_blocks_note": "computed as b = len(texts) // 3 at load time (experiments/trainval.py:256), so the blocks are NOT the 800/1600/2400 of the n=2400 corpora. Recorded explicitly because this corpus is shorter than every other entry here.",
|
| 186 |
+
"DERIVED": true,
|
| 187 |
+
"hf": null,
|
| 188 |
+
"hf_note": "INTENTIONALLY null. This corpus has NO independent upstream. It is produced deterministically from OOD_News_2024 by the rule below and must never be given a fabricated dataset revision. To regenerate it, rebuild the parent at the revision pinned in corpora/dataset_revisions.json and re-apply the rule.",
|
| 189 |
+
"derived_from": {
|
| 190 |
+
"parent": "OOD_News_2024",
|
| 191 |
+
"parent_file": "OOD_News_2024.jsonl",
|
| 192 |
+
"parent_sha256": "45aaab463d8987725fe9b7edb5bd316a29c2d547a5a1ffd2c54f41df3be1df80",
|
| 193 |
+
"parent_n": 2400,
|
| 194 |
+
"parent_upstream": {
|
| 195 |
+
"path": "RealTimeData/bbc_news_alltime",
|
| 196 |
+
"name": "2024-06",
|
| 197 |
+
"split": "train",
|
| 198 |
+
"sha": "8dd1ecdc92ac43f9c04a3da3e945537dbb08179b"
|
| 199 |
+
},
|
| 200 |
+
"rule": "exact-text deduplication: keep the FIRST occurrence of each distinct `text` value, preserving the parent's record order; drop every later exact duplicate",
|
| 201 |
+
"rule_verified": "verified 2026-09-01 by recomputation: first-occurrence dedup of the parent reproduces this file's text sequence exactly (order-sensitive equality True); sorted(distinct) does NOT (False), so the order is first-occurrence, not sorted",
|
| 202 |
+
"records_in": 2400,
|
| 203 |
+
"records_out": 1598,
|
| 204 |
+
"exact_duplicates_removed": 802,
|
| 205 |
+
"text_field": "text",
|
| 206 |
+
"is_strict_subset_of_parent": true
|
| 207 |
+
},
|
| 208 |
+
"claim": "DERIVED CONTROL. The 2024-06 BBC News panel with its 802 exact-duplicate records removed. It exists because the parent contains syndicated/repeated articles, which inflate any recurrence-based statistic. It is NOT an independent OOD rung: it shares the parent's upstream, register and time window entirely.",
|
| 209 |
+
"status": "DERIVED \u2014 provenance is the parent's plus the recorded rule. Excluded BY DESIGN from the E48b containment index (experiments/t48b_exposure_growth.py:77-81 --exclude default): it was created after the k-gram shards were built and is a strict subset of OOD_News_2024, so indexing it would double-count the parent's k-grams.",
|
| 210 |
+
"used_by": [
|
| 211 |
+
"results/e48/lens_OOD_News_2024_dedup_410m_n200_s{0,1,2}.pt",
|
| 212 |
+
"results/e48/tv_OOD_News_2024_dedup_s{0,1,2}.json"
|
| 213 |
+
],
|
| 214 |
+
"scanned": null,
|
| 215 |
+
"recorded_on": "2026-09-01",
|
| 216 |
+
"recorded_why": "this corpus had operators fitted on it but NO entry in this manifest and NO row in corpora/dataset_revisions.json, whose README claims to be the source of truth for every corpus build. Gap closed by recording the derivation, not by inventing an upstream pin."
|
| 217 |
+
}
|
| 218 |
+
}
|
results/ctrlbf1_control_backfill.json
ADDED
|
@@ -0,0 +1,305 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CTRLBF1 \u2014 backfill of three already-completed, previously unrecorded controls",
|
| 3 |
+
"prereg": [
|
| 4 |
+
"docs/experiments/preregs/superseded/PREREG_E36_QLADDER.md (section E35)",
|
| 5 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md"
|
| 6 |
+
],
|
| 7 |
+
"status": "RECORDED FROM EXISTING OUTPUTS \u2014 no new experiment, no new threshold",
|
| 8 |
+
"recomputes_not_remeasures": true,
|
| 9 |
+
"why": "CLAUDE.md rule 10 requires a control to be named with the number it produced. These three were satisfied by data already on disk but were recorded as 'NOT RECORDED' (E35 M1-C3, M1-C4) or '(full-grid arms outstanding)' (P0 C2).",
|
| 10 |
+
"E35_M1_C3": {
|
| 11 |
+
"requirement_verbatim": "post-2020 text | must score below M1-C1. If not, either the date argument or the index is wrong",
|
| 12 |
+
"m1_c1_operationalised": "M1-C1 is 'a known Pile component ... high containment'. The bar used is the LOWEST of the five in-stream Pile components at k=32, 20/20 coverage \u2014 the most demanding reading of 'below M1-C1'.",
|
| 13 |
+
"k": 32,
|
| 14 |
+
"coverage": "20/20",
|
| 15 |
+
"in_stream_final": {
|
| 16 |
+
"Pile-CC": 0.9276060773863086,
|
| 17 |
+
"Github": 0.9340004374453194,
|
| 18 |
+
"StackExchange": 0.9179914195561899,
|
| 19 |
+
"USPTO_Backgrounds": 0.9267616923398619,
|
| 20 |
+
"Wikipedia_en": 0.9215951727072293
|
| 21 |
+
},
|
| 22 |
+
"lowest_in_stream": 0.9179914195561899,
|
| 23 |
+
"post_2020": {
|
| 24 |
+
"CONTROL_PubMed_2023": {
|
| 25 |
+
"containment_20of20": 0.7927273508312213,
|
| 26 |
+
"below_lowest_in_stream": true
|
| 27 |
+
},
|
| 28 |
+
"OOD_Wikipedia_2023": {
|
| 29 |
+
"containment_20of20": 0.26957534299550123,
|
| 30 |
+
"below_lowest_in_stream": true
|
| 31 |
+
},
|
| 32 |
+
"OOD_arXiv_2023": {
|
| 33 |
+
"containment_20of20": 8.598932872430531e-05,
|
| 34 |
+
"below_lowest_in_stream": true
|
| 35 |
+
},
|
| 36 |
+
"OOD_News_2024": {
|
| 37 |
+
"containment_20of20": 0.0007487928942136383,
|
| 38 |
+
"below_lowest_in_stream": true
|
| 39 |
+
}
|
| 40 |
+
},
|
| 41 |
+
"fires": true,
|
| 42 |
+
"source": "results/e48b_exposure_growth.json -> containment_by_coverage"
|
| 43 |
+
},
|
| 44 |
+
"E35_M1_C4": {
|
| 45 |
+
"requirement_verbatim": "one shard vs twenty | containment on 1/20 of the stream must be a strict lower bound on containment over 20/20",
|
| 46 |
+
"k": 32,
|
| 47 |
+
"by_corpus": {
|
| 48 |
+
"CONTROL_PubMed_2023": {
|
| 49 |
+
"containment_1of20": 0.061166716783366985,
|
| 50 |
+
"containment_20of20": 0.7927273508312213,
|
| 51 |
+
"is_lower_bound": true,
|
| 52 |
+
"monotone_nondecreasing": true,
|
| 53 |
+
"n_coverage_points": 20
|
| 54 |
+
},
|
| 55 |
+
"Github": {
|
| 56 |
+
"containment_1of20": 0.3742096158434741,
|
| 57 |
+
"containment_20of20": 0.9340004374453194,
|
| 58 |
+
"is_lower_bound": true,
|
| 59 |
+
"monotone_nondecreasing": true,
|
| 60 |
+
"n_coverage_points": 20
|
| 61 |
+
},
|
| 62 |
+
"OOD_CommonPile": {
|
| 63 |
+
"containment_1of20": 0.0,
|
| 64 |
+
"containment_20of20": 0.00125,
|
| 65 |
+
"is_lower_bound": true,
|
| 66 |
+
"monotone_nondecreasing": true,
|
| 67 |
+
"n_coverage_points": 20
|
| 68 |
+
},
|
| 69 |
+
"OOD_News_2024": {
|
| 70 |
+
"containment_1of20": 1.936533347104237e-05,
|
| 71 |
+
"containment_20of20": 0.0007487928942136383,
|
| 72 |
+
"is_lower_bound": true,
|
| 73 |
+
"monotone_nondecreasing": true,
|
| 74 |
+
"n_coverage_points": 20
|
| 75 |
+
},
|
| 76 |
+
"OOD_Wikipedia_2023": {
|
| 77 |
+
"containment_1of20": 0.022133038856352236,
|
| 78 |
+
"containment_20of20": 0.26957534299550123,
|
| 79 |
+
"is_lower_bound": true,
|
| 80 |
+
"monotone_nondecreasing": true,
|
| 81 |
+
"n_coverage_points": 20
|
| 82 |
+
},
|
| 83 |
+
"OOD_arXiv_2023": {
|
| 84 |
+
"containment_1of20": 0.0,
|
| 85 |
+
"containment_20of20": 8.598932872430531e-05,
|
| 86 |
+
"is_lower_bound": true,
|
| 87 |
+
"monotone_nondecreasing": true,
|
| 88 |
+
"n_coverage_points": 20
|
| 89 |
+
},
|
| 90 |
+
"Pile-CC": {
|
| 91 |
+
"containment_1of20": 0.09955354052744514,
|
| 92 |
+
"containment_20of20": 0.9276060773863086,
|
| 93 |
+
"is_lower_bound": true,
|
| 94 |
+
"monotone_nondecreasing": true,
|
| 95 |
+
"n_coverage_points": 20
|
| 96 |
+
},
|
| 97 |
+
"StackExchange": {
|
| 98 |
+
"containment_1of20": 0.07686719226900293,
|
| 99 |
+
"containment_20of20": 0.9179914195561899,
|
| 100 |
+
"is_lower_bound": true,
|
| 101 |
+
"monotone_nondecreasing": true,
|
| 102 |
+
"n_coverage_points": 20
|
| 103 |
+
},
|
| 104 |
+
"TRAP_FineWeb": {
|
| 105 |
+
"containment_1of20": 0.007237811980212063,
|
| 106 |
+
"containment_20of20": 0.06309446548070848,
|
| 107 |
+
"is_lower_bound": true,
|
| 108 |
+
"monotone_nondecreasing": true,
|
| 109 |
+
"n_coverage_points": 20
|
| 110 |
+
},
|
| 111 |
+
"USPTO_Backgrounds": {
|
| 112 |
+
"containment_1of20": 0.12033439046490556,
|
| 113 |
+
"containment_20of20": 0.9267616923398619,
|
| 114 |
+
"is_lower_bound": true,
|
| 115 |
+
"monotone_nondecreasing": true,
|
| 116 |
+
"n_coverage_points": 20
|
| 117 |
+
},
|
| 118 |
+
"Wikipedia_en": {
|
| 119 |
+
"containment_1of20": 0.07066448338254318,
|
| 120 |
+
"containment_20of20": 0.9215951727072293,
|
| 121 |
+
"is_lower_bound": true,
|
| 122 |
+
"monotone_nondecreasing": true,
|
| 123 |
+
"n_coverage_points": 20
|
| 124 |
+
},
|
| 125 |
+
"_CONTROL_random": {
|
| 126 |
+
"containment_1of20": 0.0,
|
| 127 |
+
"containment_20of20": 0.0,
|
| 128 |
+
"is_lower_bound": true,
|
| 129 |
+
"monotone_nondecreasing": true,
|
| 130 |
+
"n_coverage_points": 20
|
| 131 |
+
}
|
| 132 |
+
},
|
| 133 |
+
"strict_increase_for_in_stream_corpora": {
|
| 134 |
+
"Pile-CC": true,
|
| 135 |
+
"Github": true,
|
| 136 |
+
"StackExchange": true,
|
| 137 |
+
"USPTO_Backgrounds": true,
|
| 138 |
+
"Wikipedia_en": true
|
| 139 |
+
},
|
| 140 |
+
"fires": true,
|
| 141 |
+
"note": "'strict' is checked as strict only for corpora actually present in the stream; a genuinely absent corpus pinned at 0.0 at every coverage satisfies the lower-bound requirement with equality, which is the intended behaviour and is what E48b's growth-rate reading rests on.",
|
| 142 |
+
"source": "results/e48b_exposure_growth.json -> containment_by_coverage"
|
| 143 |
+
},
|
| 144 |
+
"P0_C2": {
|
| 145 |
+
"requirement_verbatim": "C2 - worker-count invariance. --workers 1, 2 and 8 must give identical output. Required: max_abs_diff = 0.0 between every pair, on the same 128 values.",
|
| 146 |
+
"arms_found": {
|
| 147 |
+
"w1": {
|
| 148 |
+
"file": "e52_factorial_410m_rstrip_w1.json",
|
| 149 |
+
"workers": 1,
|
| 150 |
+
"D": 0.000845372218579525
|
| 151 |
+
},
|
| 152 |
+
"w2": {
|
| 153 |
+
"file": "e52_factorial_410m_rstrip_w2.json",
|
| 154 |
+
"workers": 2,
|
| 155 |
+
"D": 0.000845372218579525
|
| 156 |
+
},
|
| 157 |
+
"w8": {
|
| 158 |
+
"file": "e52_factorial_410m_rstrip_w8.json",
|
| 159 |
+
"workers": 8,
|
| 160 |
+
"D": 0.000845372218579525
|
| 161 |
+
},
|
| 162 |
+
"w6": {
|
| 163 |
+
"file": "e52_factorial_410m_rstrip.json",
|
| 164 |
+
"workers": 6,
|
| 165 |
+
"D": 0.000845372218579525
|
| 166 |
+
}
|
| 167 |
+
},
|
| 168 |
+
"n_values_compared": 128,
|
| 169 |
+
"same_key_set_across_arms": true,
|
| 170 |
+
"pairwise_max_abs_diff": {
|
| 171 |
+
"w1_vs_w2": 0.0,
|
| 172 |
+
"w1_vs_w6": 0.0,
|
| 173 |
+
"w1_vs_w8": 0.0,
|
| 174 |
+
"w2_vs_w6": 0.0,
|
| 175 |
+
"w2_vs_w8": 0.0,
|
| 176 |
+
"w6_vs_w8": 0.0
|
| 177 |
+
},
|
| 178 |
+
"pairwise_max_abs_diff_registered_arms_only": {
|
| 179 |
+
"w1_vs_w2": 0.0,
|
| 180 |
+
"w1_vs_w8": 0.0,
|
| 181 |
+
"w2_vs_w8": 0.0
|
| 182 |
+
},
|
| 183 |
+
"max_abs_diff_registered": 0.0,
|
| 184 |
+
"max_abs_diff_including_w6_reference": 0.0,
|
| 185 |
+
"fires": true,
|
| 186 |
+
"note": "The 128 values are the 64 fit x read cells under each of the two aggregations. w6 is the reference run (--workers 6) and is reported alongside but is not one of the registered 1/2/8 arms.",
|
| 187 |
+
"source": "results/e52_factorial_410m_rstrip{,_w1,_w2,_w8}.json -> by_aggregation.{persist,min}.matrix"
|
| 188 |
+
},
|
| 189 |
+
"controls_fired": {
|
| 190 |
+
"E35_M1_C3": true,
|
| 191 |
+
"E35_M1_C4": true,
|
| 192 |
+
"P0_C2": true
|
| 193 |
+
},
|
| 194 |
+
"VERDICT": "ALL THREE RECORDED. E35 M1-C3 FIRES: all 4 post-2020 panels sit below the lowest in-stream Pile component (0.9180) at k=32, 20/20 \u2014 CONTROL_PubMed_2023 0.79273, OOD_Wikipedia_2023 0.26958, OOD_arXiv_2023 0.00009, OOD_News_2024 0.00075. E35 M1-C4 FIRES: containment at 1/20 coverage is a lower bound on 20/20 for 12/12 corpora and the series is monotone non-decreasing in all of them. P0 C2 FIRES: max_abs_diff = 0.0 across every pair of --workers 1/2/8 on all 128 values. No threshold was introduced here; each requirement is quoted from its own registration.",
|
| 195 |
+
"provenance": {
|
| 196 |
+
"schema_version": 1,
|
| 197 |
+
"experiment": "CTRLBF1 \u2014 backfill of three already-completed, previously unrecorded controls",
|
| 198 |
+
"utc": "2026-09-01T20:35:05+00:00",
|
| 199 |
+
"script": {
|
| 200 |
+
"path": "experiments/ctrlbf1_control_backfill.py",
|
| 201 |
+
"sha256": "6b944b4756c5d93fc012771d2bf57553b52192cc5832ee7bff360489515fdfb4",
|
| 202 |
+
"n_lines": 156
|
| 203 |
+
},
|
| 204 |
+
"git": {
|
| 205 |
+
"commit": "9987d82887ed339bf69b5e7fa1e4fc98d5bbb4ab",
|
| 206 |
+
"commit_short": "9987d82887ed",
|
| 207 |
+
"git_repository_present": true,
|
| 208 |
+
"branch": "pythia",
|
| 209 |
+
"dirty": true,
|
| 210 |
+
"dirty_files": [
|
| 211 |
+
"CONTEXT.md",
|
| 212 |
+
"NEELNANDA_REP.md",
|
| 213 |
+
"corpora/dataset_revisions.json",
|
| 214 |
+
"corpora/dataset_revisions.json.bak",
|
| 215 |
+
"corpora/manifest_ood.json",
|
| 216 |
+
"corpora/manifest_ood.json.bak",
|
| 217 |
+
"docs/experiments/COVERAGE.md",
|
| 218 |
+
"docs/experiments/METRICS.md",
|
| 219 |
+
"docs/experiments/descriptions/D1.md",
|
| 220 |
+
"docs/experiments/descriptions/E33.md",
|
| 221 |
+
"docs/experiments/descriptions/E34A.md",
|
| 222 |
+
"docs/experiments/descriptions/E34B.md",
|
| 223 |
+
"docs/experiments/descriptions/E36.md",
|
| 224 |
+
"docs/experiments/descriptions/E37.md",
|
| 225 |
+
"docs/experiments/descriptions/E44.md",
|
| 226 |
+
"docs/experiments/descriptions/E58.md",
|
| 227 |
+
"docs/experiments/preregs/E65R_phase1_rstrip_rescore.md",
|
| 228 |
+
"docs/experiments/preregs/E65T_threshold_defect_adjudication.md",
|
| 229 |
+
"docs/experiments/preregs/E66d_s_quantization_limited.md",
|
| 230 |
+
"docs/experiments/preregs/EFFRANK1_effective_rank_adjudication.md",
|
| 231 |
+
"docs/experiments/preregs/superseded/PREREG_PYTHIA_T7_v2.md",
|
| 232 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 233 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 234 |
+
"docs/reproducibility/SCRIPT_PROVENANCE.md",
|
| 235 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 236 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 237 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 238 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 239 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/AGENT_REPORT.md",
|
| 240 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/CONTEXT_UNDER_REVIEW.md",
|
| 241 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 242 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/JUSTIFIABILITY.md",
|
| 243 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/LEDGER.md",
|
| 244 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md",
|
| 245 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/SUSPICIONS.md",
|
| 246 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Github_s0.json",
|
| 247 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Github_s1.json",
|
| 248 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Github_s2.json",
|
| 249 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Pile-CC_s0.json",
|
| 250 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Pile-CC_s1.json",
|
| 251 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Pile-CC_s2.json",
|
| 252 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_StackExchange_s0.json",
|
| 253 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_StackExchange_s1.json",
|
| 254 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_StackExchange_s2.json",
|
| 255 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_USPTO_Backgrounds_s0.json",
|
| 256 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_USPTO_Backgrounds_s1.json",
|
| 257 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_USPTO_Backgrounds_s2.json",
|
| 258 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Wikipedia_en_s0.json",
|
| 259 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Wikipedia_en_s1.json",
|
| 260 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/data/rstrip_rescore/rs_Wikipedia_en_s2.json",
|
| 261 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/balance_test.py",
|
| 262 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/check_payload.py",
|
| 263 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/coverage2.py",
|
| 264 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/derange.py",
|
| 265 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/derangement_result.json",
|
| 266 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/logit_base.py",
|
| 267 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/logit_baseline.json",
|
| 268 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/my_anova.py",
|
| 269 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/my_e52.py",
|
| 270 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/my_rescore.py",
|
| 271 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/prereg2.py",
|
| 272 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/prereg_audit.py",
|
| 273 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/rescore_all.py",
|
| 274 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/scripts/rstrip_anova.py"
|
| 275 |
+
],
|
| 276 |
+
"dirty_diff_sha256": "719949b81d9871a26c26430713cafa9ad3b949cc5eea0ad080ae38307d75f63d",
|
| 277 |
+
"describe": "9987d82-dirty"
|
| 278 |
+
},
|
| 279 |
+
"argv": [
|
| 280 |
+
"experiments/ctrlbf1_control_backfill.py"
|
| 281 |
+
],
|
| 282 |
+
"cwd": ".",
|
| 283 |
+
"inputs": [],
|
| 284 |
+
"env": {
|
| 285 |
+
"python": "3.14.2",
|
| 286 |
+
"torch": "2.13.0",
|
| 287 |
+
"transformers": "5.14.1",
|
| 288 |
+
"numpy": "2.5.1",
|
| 289 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 290 |
+
"machine": "arm64",
|
| 291 |
+
"tf32": {
|
| 292 |
+
"allow_tf32_matmul": false,
|
| 293 |
+
"allow_tf32_cudnn": true,
|
| 294 |
+
"NVIDIA_TF32_OVERRIDE": null
|
| 295 |
+
},
|
| 296 |
+
"threads": {
|
| 297 |
+
"torch_num_threads": 10,
|
| 298 |
+
"torch_num_interop_threads": 14,
|
| 299 |
+
"OMP_NUM_THREADS": null,
|
| 300 |
+
"os_cpu_count": 14
|
| 301 |
+
}
|
| 302 |
+
},
|
| 303 |
+
"payload_sha256": "244eafca442da2a20faa694df90c22fe3bdd41e299594c22382a37b5825e4962"
|
| 304 |
+
}
|
| 305 |
+
}
|
results/cv1_answer_competence.json
ADDED
|
@@ -0,0 +1,288 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV1 \u2014 answer competence: can the model do the task at all?",
|
| 3 |
+
"status": "MEASUREMENT ONLY \u2014 no registered decision rule",
|
| 4 |
+
"why": "existing capability_* fields measure the INTERMEDIATE's rank, not the ANSWER's; t36_qladder.py:483-489 records the registered competence gate as degenerate and unable to fire. This supplies the missing number.",
|
| 5 |
+
"readout_convention": "STRIPPED (corrected) \u2014 rank of the target's first token at the final position of the rstripped prompt",
|
| 6 |
+
"sets_with_target": [
|
| 7 |
+
"multihop",
|
| 8 |
+
"order-ops",
|
| 9 |
+
"multilingual"
|
| 10 |
+
],
|
| 11 |
+
"sets_without_target": [
|
| 12 |
+
"poetry",
|
| 13 |
+
"typo",
|
| 14 |
+
"association"
|
| 15 |
+
],
|
| 16 |
+
"note_on_undefined": "poetry/typo/association carry target: null. The annotated intermediate is the object of interest and there is no separate answer, so answer competence is UNDEFINED for them \u2014 not zero.",
|
| 17 |
+
"by_model": {
|
| 18 |
+
"70m": {
|
| 19 |
+
"multihop": {
|
| 20 |
+
"n_scored": 93,
|
| 21 |
+
"n_multitoken_answers_scored_on_first_token": 3,
|
| 22 |
+
"top1_rate": 0.010752688172043012,
|
| 23 |
+
"top10_rate": 0.13978494623655913,
|
| 24 |
+
"median_rank": 80,
|
| 25 |
+
"mean_rank": 1106.5806451612902,
|
| 26 |
+
"median_prob": 0.0015416988171637058
|
| 27 |
+
},
|
| 28 |
+
"order-ops": {
|
| 29 |
+
"n_scored": 55,
|
| 30 |
+
"n_multitoken_answers_scored_on_first_token": 0,
|
| 31 |
+
"top1_rate": 0.03636363636363636,
|
| 32 |
+
"top10_rate": 0.38181818181818183,
|
| 33 |
+
"median_rank": 18,
|
| 34 |
+
"mean_rank": 22.09090909090909,
|
| 35 |
+
"median_prob": 0.006405120715498924
|
| 36 |
+
},
|
| 37 |
+
"multilingual": {
|
| 38 |
+
"n_scored": 107,
|
| 39 |
+
"n_multitoken_answers_scored_on_first_token": 92,
|
| 40 |
+
"top1_rate": 0.009345794392523364,
|
| 41 |
+
"top10_rate": 0.08411214953271028,
|
| 42 |
+
"median_rank": 3436,
|
| 43 |
+
"mean_rank": 4615.915887850468,
|
| 44 |
+
"median_prob": 2.3302003683056682e-05
|
| 45 |
+
},
|
| 46 |
+
"_pooled": {
|
| 47 |
+
"n": 255,
|
| 48 |
+
"top1_rate": 0.01568627450980392,
|
| 49 |
+
"top10_rate": 0.16862745098039217
|
| 50 |
+
}
|
| 51 |
+
},
|
| 52 |
+
"160m": {
|
| 53 |
+
"multihop": {
|
| 54 |
+
"n_scored": 93,
|
| 55 |
+
"n_multitoken_answers_scored_on_first_token": 3,
|
| 56 |
+
"top1_rate": 0.03225806451612903,
|
| 57 |
+
"top10_rate": 0.25806451612903225,
|
| 58 |
+
"median_rank": 42,
|
| 59 |
+
"mean_rank": 628.3655913978495,
|
| 60 |
+
"median_prob": 0.0031900564208626747
|
| 61 |
+
},
|
| 62 |
+
"order-ops": {
|
| 63 |
+
"n_scored": 55,
|
| 64 |
+
"n_multitoken_answers_scored_on_first_token": 0,
|
| 65 |
+
"top1_rate": 0.0,
|
| 66 |
+
"top10_rate": 0.43636363636363634,
|
| 67 |
+
"median_rank": 13,
|
| 68 |
+
"mean_rank": 14.49090909090909,
|
| 69 |
+
"median_prob": 0.01120622269809246
|
| 70 |
+
},
|
| 71 |
+
"multilingual": {
|
| 72 |
+
"n_scored": 107,
|
| 73 |
+
"n_multitoken_answers_scored_on_first_token": 92,
|
| 74 |
+
"top1_rate": 0.009345794392523364,
|
| 75 |
+
"top10_rate": 0.08411214953271028,
|
| 76 |
+
"median_rank": 3788,
|
| 77 |
+
"mean_rank": 4630.429906542056,
|
| 78 |
+
"median_prob": 1.6619400412309915e-05
|
| 79 |
+
},
|
| 80 |
+
"_pooled": {
|
| 81 |
+
"n": 255,
|
| 82 |
+
"top1_rate": 0.01568627450980392,
|
| 83 |
+
"top10_rate": 0.2235294117647059
|
| 84 |
+
}
|
| 85 |
+
},
|
| 86 |
+
"410m": {
|
| 87 |
+
"multihop": {
|
| 88 |
+
"n_scored": 93,
|
| 89 |
+
"n_multitoken_answers_scored_on_first_token": 3,
|
| 90 |
+
"top1_rate": 0.06451612903225806,
|
| 91 |
+
"top10_rate": 0.3333333333333333,
|
| 92 |
+
"median_rank": 31,
|
| 93 |
+
"mean_rank": 460.9139784946237,
|
| 94 |
+
"median_prob": 0.005166209302842617
|
| 95 |
+
},
|
| 96 |
+
"order-ops": {
|
| 97 |
+
"n_scored": 55,
|
| 98 |
+
"n_multitoken_answers_scored_on_first_token": 0,
|
| 99 |
+
"top1_rate": 0.09090909090909091,
|
| 100 |
+
"top10_rate": 0.5454545454545454,
|
| 101 |
+
"median_rank": 9,
|
| 102 |
+
"mean_rank": 11.909090909090908,
|
| 103 |
+
"median_prob": 0.026608118787407875
|
| 104 |
+
},
|
| 105 |
+
"multilingual": {
|
| 106 |
+
"n_scored": 107,
|
| 107 |
+
"n_multitoken_answers_scored_on_first_token": 92,
|
| 108 |
+
"top1_rate": 0.009345794392523364,
|
| 109 |
+
"top10_rate": 0.056074766355140186,
|
| 110 |
+
"median_rank": 3044,
|
| 111 |
+
"mean_rank": 4031.233644859813,
|
| 112 |
+
"median_prob": 1.0294868843629956e-05
|
| 113 |
+
},
|
| 114 |
+
"_pooled": {
|
| 115 |
+
"n": 255,
|
| 116 |
+
"top1_rate": 0.047058823529411764,
|
| 117 |
+
"top10_rate": 0.2627450980392157
|
| 118 |
+
}
|
| 119 |
+
},
|
| 120 |
+
"1b": {
|
| 121 |
+
"multihop": {
|
| 122 |
+
"n_scored": 93,
|
| 123 |
+
"n_multitoken_answers_scored_on_first_token": 3,
|
| 124 |
+
"top1_rate": 0.07526881720430108,
|
| 125 |
+
"top10_rate": 0.3333333333333333,
|
| 126 |
+
"median_rank": 18,
|
| 127 |
+
"mean_rank": 316.0967741935484,
|
| 128 |
+
"median_prob": 0.011097713373601437
|
| 129 |
+
},
|
| 130 |
+
"order-ops": {
|
| 131 |
+
"n_scored": 55,
|
| 132 |
+
"n_multitoken_answers_scored_on_first_token": 0,
|
| 133 |
+
"top1_rate": 0.07272727272727272,
|
| 134 |
+
"top10_rate": 0.6727272727272727,
|
| 135 |
+
"median_rank": 7,
|
| 136 |
+
"mean_rank": 9.163636363636364,
|
| 137 |
+
"median_prob": 0.029681021347641945
|
| 138 |
+
},
|
| 139 |
+
"multilingual": {
|
| 140 |
+
"n_scored": 107,
|
| 141 |
+
"n_multitoken_answers_scored_on_first_token": 92,
|
| 142 |
+
"top1_rate": 0.0,
|
| 143 |
+
"top10_rate": 0.06542056074766354,
|
| 144 |
+
"median_rank": 1649,
|
| 145 |
+
"mean_rank": 2892.6168224299067,
|
| 146 |
+
"median_prob": 3.0693336157128215e-05
|
| 147 |
+
},
|
| 148 |
+
"_pooled": {
|
| 149 |
+
"n": 255,
|
| 150 |
+
"top1_rate": 0.043137254901960784,
|
| 151 |
+
"top10_rate": 0.29411764705882354
|
| 152 |
+
}
|
| 153 |
+
}
|
| 154 |
+
},
|
| 155 |
+
"VERDICT": "Answer competence on the three sets carrying a target. 70m: top1 1.6%, top10 16.9% (n=255 items with a target) | 160m: top1 1.6%, top10 22.4% (n=255 items with a target) | 410m: top1 4.7%, top10 26.3% (n=255 items with a target) | 1b: top1 4.3%, top10 29.4% (n=255 items with a target)",
|
| 156 |
+
"provenance": {
|
| 157 |
+
"schema_version": 1,
|
| 158 |
+
"experiment": "CV1",
|
| 159 |
+
"utc": "2026-08-30T06:32:53+00:00",
|
| 160 |
+
"script": {
|
| 161 |
+
"path": "experiments/cv1_answer_competence.py",
|
| 162 |
+
"sha256": "29279dcb364dec36ae9f32438ab9ac607cc8b89983cf20c5796662e9a56a7a21",
|
| 163 |
+
"n_lines": 138
|
| 164 |
+
},
|
| 165 |
+
"git": {
|
| 166 |
+
"commit": "c9c49697e295c0fec4225552836fc9733b279f75",
|
| 167 |
+
"commit_short": "c9c49697e295",
|
| 168 |
+
"git_repository_present": true,
|
| 169 |
+
"branch": "pythia",
|
| 170 |
+
"dirty": true,
|
| 171 |
+
"dirty_files": [
|
| 172 |
+
"CLAUDE.md",
|
| 173 |
+
"README.md",
|
| 174 |
+
"corpora/manifest_ood.json",
|
| 175 |
+
"docs/HANDOFF.md",
|
| 176 |
+
"docs/REPO_REVIEW_PROMPT.md",
|
| 177 |
+
"docs/context/AGGREGATION_POLICY.md",
|
| 178 |
+
"docs/context/CONTEXT.md",
|
| 179 |
+
"docs/context/IMPLEMENTATION_RULES.md",
|
| 180 |
+
"docs/context/STATE.md",
|
| 181 |
+
"docs/experiments/INDEX.md",
|
| 182 |
+
"docs/experiments/METRICS.md",
|
| 183 |
+
"docs/experiments/PROPOSED.md",
|
| 184 |
+
"docs/experiments/PROVENANCE.md",
|
| 185 |
+
"docs/experiments/descriptions/E64.md",
|
| 186 |
+
"docs/experiments/descriptions/EN1.md",
|
| 187 |
+
"docs/experiments/preregs/CV3_margins.md",
|
| 188 |
+
"docs/experiments/preregs/CV4_capability_ladder.md",
|
| 189 |
+
"docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 190 |
+
"docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 191 |
+
"docs/experiments/preregs/CV7_1b_rung.md",
|
| 192 |
+
"docs/experiments/preregs/DN1_design_nulls.md",
|
| 193 |
+
"docs/experiments/preregs/E33_logit_baseline.md",
|
| 194 |
+
"docs/experiments/preregs/E36_qladder.md",
|
| 195 |
+
"docs/experiments/preregs/E48_crossover.md",
|
| 196 |
+
"docs/experiments/preregs/E48b_containment.md",
|
| 197 |
+
"docs/experiments/preregs/E48c_exposure_vs_read.md",
|
| 198 |
+
"docs/experiments/preregs/E51_interaction.md",
|
| 199 |
+
"docs/experiments/preregs/E52_factorial.md",
|
| 200 |
+
"docs/experiments/preregs/E53_ladder_summary.md",
|
| 201 |
+
"docs/experiments/preregs/E55_matrix_robustness.md",
|
| 202 |
+
"docs/experiments/preregs/E56_predictor_registry.md",
|
| 203 |
+
"docs/experiments/preregs/E57_grid_variance_ci.md",
|
| 204 |
+
"docs/experiments/preregs/E58_algebra_audit.md",
|
| 205 |
+
"docs/experiments/preregs/E59_read_dose_response.md",
|
| 206 |
+
"docs/experiments/preregs/E60_fitter_determinism.md",
|
| 207 |
+
"docs/experiments/preregs/E61_randomized_network_null.md",
|
| 208 |
+
"docs/experiments/preregs/E62_ladder1b_corrected_band.md",
|
| 209 |
+
"docs/experiments/preregs/E65_training_axis_floor.md",
|
| 210 |
+
"docs/experiments/preregs/E66b_determinism_floor.md",
|
| 211 |
+
"docs/experiments/preregs/OS1_operator_space.md",
|
| 212 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md",
|
| 213 |
+
"docs/experiments/preregs/R1_grid_rstrip.md",
|
| 214 |
+
"docs/experiments/preregs/R2_e48_rstrip_arm.md",
|
| 215 |
+
"docs/experiments/preregs/R3_ladder410_cpu.md",
|
| 216 |
+
"docs/experiments/preregs/R4_corrections.md",
|
| 217 |
+
"docs/experiments/preregs/R4b_e36_rstrip.md",
|
| 218 |
+
"docs/experiments/preregs/R5_corpus_axis_uncertainty.md",
|
| 219 |
+
"docs/experiments/preregs/R6_within_source_resampling.md",
|
| 220 |
+
"docs/experiments/preregs/R8_ladder_rstrip_S2.md",
|
| 221 |
+
"docs/experiments/preregs/R9_permutation_calibrated_min.md",
|
| 222 |
+
"docs/experiments/preregs/superseded/PREREG_E48_CROSSOVER.md",
|
| 223 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 224 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 225 |
+
"docs/reproducibility/RELEASE.md",
|
| 226 |
+
"docs/reproducibility/REVIEWER_REPRO.md",
|
| 227 |
+
"docs/reproducibility/T1_FINDING_E33_BAND.md",
|
| 228 |
+
"docs/reproducibility/T1_SWEEP_PREDICTIONS.md",
|
| 229 |
+
"docs/reproducibility/T1_SWEEP_SPEC.md",
|
| 230 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 231 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 232 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 233 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 234 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 235 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md"
|
| 236 |
+
],
|
| 237 |
+
"dirty_diff_sha256": "ec61dde1905616f407f51d5f6fe3ea37c9c6e3e1ade6998cd0644527d080c600",
|
| 238 |
+
"describe": "c9c4969-dirty"
|
| 239 |
+
},
|
| 240 |
+
"argv": [
|
| 241 |
+
"experiments/cv1_answer_competence.py",
|
| 242 |
+
"--models",
|
| 243 |
+
"70m,160m,410m,1b"
|
| 244 |
+
],
|
| 245 |
+
"cwd": ".",
|
| 246 |
+
"inputs": [
|
| 247 |
+
{
|
| 248 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-multihop.json",
|
| 249 |
+
"sha256": "50b7e4c9255291c0ca2a8e94615be9f44531fa57bb1a844e4f9616056d987416",
|
| 250 |
+
"bytes": 21869,
|
| 251 |
+
"exists": true
|
| 252 |
+
},
|
| 253 |
+
{
|
| 254 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-order-ops.json",
|
| 255 |
+
"sha256": "b203206d16ff628152cc86f3838604e06cb54776f3e14fa1c34f150db8bc7560",
|
| 256 |
+
"bytes": 9589,
|
| 257 |
+
"exists": true
|
| 258 |
+
},
|
| 259 |
+
{
|
| 260 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-multilingual.json",
|
| 261 |
+
"sha256": "fa70b9bd89416a6d8d985a80dc628b109ae6fd3b25b9275c0fc5065d7ff4a0ef",
|
| 262 |
+
"bytes": 24284,
|
| 263 |
+
"exists": true
|
| 264 |
+
}
|
| 265 |
+
],
|
| 266 |
+
"env": {
|
| 267 |
+
"python": "3.14.2",
|
| 268 |
+
"torch": "2.13.0",
|
| 269 |
+
"transformers": "5.14.1",
|
| 270 |
+
"numpy": "2.5.1",
|
| 271 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 272 |
+
"machine": "arm64",
|
| 273 |
+
"hostname": "<redacted for double-blind review>",
|
| 274 |
+
"tf32": {
|
| 275 |
+
"allow_tf32_matmul": false,
|
| 276 |
+
"allow_tf32_cudnn": true,
|
| 277 |
+
"NVIDIA_TF32_OVERRIDE": null
|
| 278 |
+
},
|
| 279 |
+
"threads": {
|
| 280 |
+
"torch_num_threads": 10,
|
| 281 |
+
"torch_num_interop_threads": 14,
|
| 282 |
+
"OMP_NUM_THREADS": null,
|
| 283 |
+
"os_cpu_count": 14
|
| 284 |
+
}
|
| 285 |
+
},
|
| 286 |
+
"payload_sha256": "8809019f865fe0939cee28bd5d201f7a842b85b9cb5929440541804f3786e74d"
|
| 287 |
+
}
|
| 288 |
+
}
|
results/cv2_position_support.json
ADDED
|
@@ -0,0 +1,299 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV2 \u2014 positional support audit",
|
| 3 |
+
"status": "MEASUREMENT ONLY \u2014 no registered decision rule; this establishes or refutes the mismatch flagged in docs/validity/CONSTRUCT_VALIDITY.md section 4.2",
|
| 4 |
+
"model": "EleutherAI/pythia-410m-deduped",
|
| 5 |
+
"fitting_support": {
|
| 6 |
+
"source": "src/fastfit.py:116 -> jlens.fitting.valid_position_mask(seq_len, skip_first)",
|
| 7 |
+
"skip_first": 16,
|
| 8 |
+
"max_seq_len": 128,
|
| 9 |
+
"excludes": "first skip_first positions (attention sink) and the final position (no next-token target)",
|
| 10 |
+
"inclusive_range": [
|
| 11 |
+
16,
|
| 12 |
+
126
|
| 13 |
+
]
|
| 14 |
+
},
|
| 15 |
+
"readout_convention": "STRIPPED (corrected). Position computed on the item alone, then shifted by the prefix length \u2014 matches t36_qladder.py:27-31",
|
| 16 |
+
"by_set": {
|
| 17 |
+
"association": {
|
| 18 |
+
"n_items": 102,
|
| 19 |
+
"readout_rule": "last",
|
| 20 |
+
"median_prompt_tokens": 28,
|
| 21 |
+
"Q0": {
|
| 22 |
+
"median_read_pos": 28,
|
| 23 |
+
"min": 19,
|
| 24 |
+
"max": 49,
|
| 25 |
+
"n_in_support": 102,
|
| 26 |
+
"frac_in_support": 1.0
|
| 27 |
+
},
|
| 28 |
+
"prefixed": {
|
| 29 |
+
"median_read_pos": 156,
|
| 30 |
+
"min": 147,
|
| 31 |
+
"max": 177,
|
| 32 |
+
"n_in_support": 0,
|
| 33 |
+
"frac_in_support": 0.0
|
| 34 |
+
}
|
| 35 |
+
},
|
| 36 |
+
"multihop": {
|
| 37 |
+
"n_items": 93,
|
| 38 |
+
"readout_rule": "last",
|
| 39 |
+
"median_prompt_tokens": 15,
|
| 40 |
+
"Q0": {
|
| 41 |
+
"median_read_pos": 15,
|
| 42 |
+
"min": 8,
|
| 43 |
+
"max": 40,
|
| 44 |
+
"n_in_support": 45,
|
| 45 |
+
"frac_in_support": 0.4838709677419355
|
| 46 |
+
},
|
| 47 |
+
"prefixed": {
|
| 48 |
+
"median_read_pos": 143,
|
| 49 |
+
"min": 136,
|
| 50 |
+
"max": 168,
|
| 51 |
+
"n_in_support": 0,
|
| 52 |
+
"frac_in_support": 0.0
|
| 53 |
+
}
|
| 54 |
+
},
|
| 55 |
+
"multilingual": {
|
| 56 |
+
"n_items": 107,
|
| 57 |
+
"readout_rule": "last",
|
| 58 |
+
"median_prompt_tokens": 10,
|
| 59 |
+
"Q0": {
|
| 60 |
+
"median_read_pos": 10,
|
| 61 |
+
"min": 5,
|
| 62 |
+
"max": 26,
|
| 63 |
+
"n_in_support": 6,
|
| 64 |
+
"frac_in_support": 0.056074766355140186
|
| 65 |
+
},
|
| 66 |
+
"prefixed": {
|
| 67 |
+
"median_read_pos": 138,
|
| 68 |
+
"min": 133,
|
| 69 |
+
"max": 154,
|
| 70 |
+
"n_in_support": 0,
|
| 71 |
+
"frac_in_support": 0.0
|
| 72 |
+
}
|
| 73 |
+
},
|
| 74 |
+
"order-ops": {
|
| 75 |
+
"n_items": 55,
|
| 76 |
+
"readout_rule": "last",
|
| 77 |
+
"median_prompt_tokens": 8,
|
| 78 |
+
"Q0": {
|
| 79 |
+
"median_read_pos": 8,
|
| 80 |
+
"min": 6,
|
| 81 |
+
"max": 18,
|
| 82 |
+
"n_in_support": 1,
|
| 83 |
+
"frac_in_support": 0.01818181818181818
|
| 84 |
+
},
|
| 85 |
+
"prefixed": {
|
| 86 |
+
"median_read_pos": 136,
|
| 87 |
+
"min": 134,
|
| 88 |
+
"max": 146,
|
| 89 |
+
"n_in_support": 0,
|
| 90 |
+
"frac_in_support": 0.0
|
| 91 |
+
}
|
| 92 |
+
},
|
| 93 |
+
"poetry": {
|
| 94 |
+
"n_items": 98,
|
| 95 |
+
"readout_rule": "last_newline",
|
| 96 |
+
"median_prompt_tokens": 26,
|
| 97 |
+
"Q0": {
|
| 98 |
+
"median_read_pos": 18,
|
| 99 |
+
"min": 16,
|
| 100 |
+
"max": 22,
|
| 101 |
+
"n_in_support": 98,
|
| 102 |
+
"frac_in_support": 1.0
|
| 103 |
+
},
|
| 104 |
+
"prefixed": {
|
| 105 |
+
"median_read_pos": 146,
|
| 106 |
+
"min": 144,
|
| 107 |
+
"max": 150,
|
| 108 |
+
"n_in_support": 0,
|
| 109 |
+
"frac_in_support": 0.0
|
| 110 |
+
}
|
| 111 |
+
},
|
| 112 |
+
"typo": {
|
| 113 |
+
"n_items": 96,
|
| 114 |
+
"readout_rule": "last",
|
| 115 |
+
"median_prompt_tokens": 8,
|
| 116 |
+
"Q0": {
|
| 117 |
+
"median_read_pos": 8,
|
| 118 |
+
"min": 6,
|
| 119 |
+
"max": 14,
|
| 120 |
+
"n_in_support": 0,
|
| 121 |
+
"frac_in_support": 0.0
|
| 122 |
+
},
|
| 123 |
+
"prefixed": {
|
| 124 |
+
"median_read_pos": 136,
|
| 125 |
+
"min": 134,
|
| 126 |
+
"max": 142,
|
| 127 |
+
"n_in_support": 0,
|
| 128 |
+
"frac_in_support": 0.0
|
| 129 |
+
}
|
| 130 |
+
}
|
| 131 |
+
},
|
| 132 |
+
"overall": {
|
| 133 |
+
"Q0": {
|
| 134 |
+
"n": 551,
|
| 135 |
+
"median_read_pos": 14,
|
| 136 |
+
"min": 5,
|
| 137 |
+
"max": 49,
|
| 138 |
+
"n_in_support": 252,
|
| 139 |
+
"frac_in_support": 0.4573502722323049
|
| 140 |
+
},
|
| 141 |
+
"prefixed": {
|
| 142 |
+
"n": 551,
|
| 143 |
+
"median_read_pos": 142,
|
| 144 |
+
"min": 133,
|
| 145 |
+
"max": 177,
|
| 146 |
+
"n_in_support": 0,
|
| 147 |
+
"frac_in_support": 0.0
|
| 148 |
+
}
|
| 149 |
+
},
|
| 150 |
+
"VERDICT": "Q0 reads inside the fitting support on 45.7% of items (median position 14); prefixed rungs on 0.0% (median 142). Fitting support is positions 16..126.",
|
| 151 |
+
"provenance": {
|
| 152 |
+
"schema_version": 1,
|
| 153 |
+
"experiment": "CV2",
|
| 154 |
+
"utc": "2026-08-30T00:19:04+00:00",
|
| 155 |
+
"script": {
|
| 156 |
+
"path": "experiments/cv2_position_support.py",
|
| 157 |
+
"sha256": "fee6460d8d930d3c9ebc138a29def2533ca5bd8fc2392daceed2b0644e20abbc",
|
| 158 |
+
"n_lines": 145
|
| 159 |
+
},
|
| 160 |
+
"git": {
|
| 161 |
+
"commit": "c9c49697e295c0fec4225552836fc9733b279f75",
|
| 162 |
+
"commit_short": "c9c49697e295",
|
| 163 |
+
"git_repository_present": true,
|
| 164 |
+
"branch": "pythia",
|
| 165 |
+
"dirty": true,
|
| 166 |
+
"dirty_files": [
|
| 167 |
+
"CLAUDE.md",
|
| 168 |
+
"README.md",
|
| 169 |
+
"corpora/manifest_ood.json",
|
| 170 |
+
"docs/HANDOFF.md",
|
| 171 |
+
"docs/REPO_REVIEW_PROMPT.md",
|
| 172 |
+
"docs/context/AGGREGATION_POLICY.md",
|
| 173 |
+
"docs/context/CONTEXT.md",
|
| 174 |
+
"docs/context/IMPLEMENTATION_RULES.md",
|
| 175 |
+
"docs/context/STATE.md",
|
| 176 |
+
"docs/experiments/INDEX.md",
|
| 177 |
+
"docs/experiments/METRICS.md",
|
| 178 |
+
"docs/experiments/PROPOSED.md",
|
| 179 |
+
"docs/experiments/PROVENANCE.md",
|
| 180 |
+
"docs/experiments/descriptions/E64.md",
|
| 181 |
+
"docs/experiments/descriptions/EN1.md",
|
| 182 |
+
"docs/experiments/preregs/CV3_margins.md",
|
| 183 |
+
"docs/experiments/preregs/CV4_capability_ladder.md",
|
| 184 |
+
"docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 185 |
+
"docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 186 |
+
"docs/experiments/preregs/CV7_1b_rung.md",
|
| 187 |
+
"docs/experiments/preregs/DN1_design_nulls.md",
|
| 188 |
+
"docs/experiments/preregs/E33_logit_baseline.md",
|
| 189 |
+
"docs/experiments/preregs/E36_qladder.md",
|
| 190 |
+
"docs/experiments/preregs/E48_crossover.md",
|
| 191 |
+
"docs/experiments/preregs/E48b_containment.md",
|
| 192 |
+
"docs/experiments/preregs/E48c_exposure_vs_read.md",
|
| 193 |
+
"docs/experiments/preregs/E51_interaction.md",
|
| 194 |
+
"docs/experiments/preregs/E52_factorial.md",
|
| 195 |
+
"docs/experiments/preregs/E53_ladder_summary.md",
|
| 196 |
+
"docs/experiments/preregs/E55_matrix_robustness.md",
|
| 197 |
+
"docs/experiments/preregs/E56_predictor_registry.md",
|
| 198 |
+
"docs/experiments/preregs/E57_grid_variance_ci.md",
|
| 199 |
+
"docs/experiments/preregs/E58_algebra_audit.md",
|
| 200 |
+
"docs/experiments/preregs/E59_read_dose_response.md",
|
| 201 |
+
"docs/experiments/preregs/E60_fitter_determinism.md",
|
| 202 |
+
"docs/experiments/preregs/E61_randomized_network_null.md",
|
| 203 |
+
"docs/experiments/preregs/E62_ladder1b_corrected_band.md",
|
| 204 |
+
"docs/experiments/preregs/E65_training_axis_floor.md",
|
| 205 |
+
"docs/experiments/preregs/E66b_determinism_floor.md",
|
| 206 |
+
"docs/experiments/preregs/OS1_operator_space.md",
|
| 207 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md",
|
| 208 |
+
"docs/experiments/preregs/R1_grid_rstrip.md",
|
| 209 |
+
"docs/experiments/preregs/R2_e48_rstrip_arm.md",
|
| 210 |
+
"docs/experiments/preregs/R3_ladder410_cpu.md",
|
| 211 |
+
"docs/experiments/preregs/R4_corrections.md",
|
| 212 |
+
"docs/experiments/preregs/R4b_e36_rstrip.md",
|
| 213 |
+
"docs/experiments/preregs/R5_corpus_axis_uncertainty.md",
|
| 214 |
+
"docs/experiments/preregs/R6_within_source_resampling.md",
|
| 215 |
+
"docs/experiments/preregs/R8_ladder_rstrip_S2.md",
|
| 216 |
+
"docs/experiments/preregs/R9_permutation_calibrated_min.md",
|
| 217 |
+
"docs/experiments/preregs/superseded/PREREG_E48_CROSSOVER.md",
|
| 218 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 219 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 220 |
+
"docs/reproducibility/RELEASE.md",
|
| 221 |
+
"docs/reproducibility/REVIEWER_REPRO.md",
|
| 222 |
+
"docs/reproducibility/T1_SWEEP_SPEC.md",
|
| 223 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 224 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 225 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 226 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 227 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 228 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md",
|
| 229 |
+
"docs/reviews/REVIEW_INSIGHTS_2026-08-25.md",
|
| 230 |
+
"docs/validity/CONSTRUCT_VALIDITY.md"
|
| 231 |
+
],
|
| 232 |
+
"dirty_diff_sha256": "6b143c292229b3d2480d1b73243675f0db5e23b27f66e4215e0a9c1947a761a8",
|
| 233 |
+
"describe": "c9c4969-dirty"
|
| 234 |
+
},
|
| 235 |
+
"argv": [
|
| 236 |
+
"experiments/cv2_position_support.py"
|
| 237 |
+
],
|
| 238 |
+
"cwd": ".",
|
| 239 |
+
"inputs": [
|
| 240 |
+
{
|
| 241 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-association.json",
|
| 242 |
+
"sha256": "d1a98cd4911b594282e74168091c77d849dae18ffe2acb5761074853f327d71c",
|
| 243 |
+
"bytes": 24228,
|
| 244 |
+
"exists": true
|
| 245 |
+
},
|
| 246 |
+
{
|
| 247 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-multihop.json",
|
| 248 |
+
"sha256": "50b7e4c9255291c0ca2a8e94615be9f44531fa57bb1a844e4f9616056d987416",
|
| 249 |
+
"bytes": 21869,
|
| 250 |
+
"exists": true
|
| 251 |
+
},
|
| 252 |
+
{
|
| 253 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-multilingual.json",
|
| 254 |
+
"sha256": "fa70b9bd89416a6d8d985a80dc628b109ae6fd3b25b9275c0fc5065d7ff4a0ef",
|
| 255 |
+
"bytes": 24284,
|
| 256 |
+
"exists": true
|
| 257 |
+
},
|
| 258 |
+
{
|
| 259 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-order-ops.json",
|
| 260 |
+
"sha256": "b203206d16ff628152cc86f3838604e06cb54776f3e14fa1c34f150db8bc7560",
|
| 261 |
+
"bytes": 9589,
|
| 262 |
+
"exists": true
|
| 263 |
+
},
|
| 264 |
+
{
|
| 265 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-poetry.json",
|
| 266 |
+
"sha256": "6aeb3415c5a5c3f3827c9efe63f006de02f5ef39a816bbac68e15e733aba60cc",
|
| 267 |
+
"bytes": 21533,
|
| 268 |
+
"exists": true
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-typo.json",
|
| 272 |
+
"sha256": "9d05e16b7234a57d0773d120a4e1c4e94fd3bc2235a8125d4200a70e60ab17aa",
|
| 273 |
+
"bytes": 14646,
|
| 274 |
+
"exists": true
|
| 275 |
+
}
|
| 276 |
+
],
|
| 277 |
+
"env": {
|
| 278 |
+
"python": "3.14.2",
|
| 279 |
+
"torch": "2.13.0",
|
| 280 |
+
"transformers": "5.14.1",
|
| 281 |
+
"numpy": "2.5.1",
|
| 282 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 283 |
+
"machine": "arm64",
|
| 284 |
+
"hostname": "<redacted for double-blind review>",
|
| 285 |
+
"tf32": {
|
| 286 |
+
"allow_tf32_matmul": false,
|
| 287 |
+
"allow_tf32_cudnn": true,
|
| 288 |
+
"NVIDIA_TF32_OVERRIDE": null
|
| 289 |
+
},
|
| 290 |
+
"threads": {
|
| 291 |
+
"torch_num_threads": 10,
|
| 292 |
+
"torch_num_interop_threads": 14,
|
| 293 |
+
"OMP_NUM_THREADS": null,
|
| 294 |
+
"os_cpu_count": 14
|
| 295 |
+
}
|
| 296 |
+
},
|
| 297 |
+
"payload_sha256": "1eb1ab9b76e77d6060876485b34291cf9e1c525a755ae0166d5f7ec98f6d0d92"
|
| 298 |
+
}
|
| 299 |
+
}
|
results/cv3_margins_410m.json
ADDED
|
@@ -0,0 +1,473 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV3 \u2014 margins: score effect or ranking artifact?",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV3_margins.md",
|
| 4 |
+
"status": "PRE-REGISTERED",
|
| 5 |
+
"decision_rule_verbatim": "ACCEPT (SCORE EFFECT) if PRIMARY >= 0.50 | REJECT (RANKING ARTIFACT) if PRIMARY < 0.25 \u2014 stop and alert, do not re-cut | UNCLEAR if 0.25 <= PRIMARY < 0.50 \u2014 report and stop",
|
| 6 |
+
"model": "EleutherAI/pythia-410m-deduped",
|
| 7 |
+
"band": [
|
| 8 |
+
9,
|
| 9 |
+
10,
|
| 10 |
+
11,
|
| 11 |
+
12,
|
| 12 |
+
13,
|
| 13 |
+
14,
|
| 14 |
+
15,
|
| 15 |
+
16,
|
| 16 |
+
17,
|
| 17 |
+
18,
|
| 18 |
+
19,
|
| 19 |
+
20,
|
| 20 |
+
21
|
| 21 |
+
],
|
| 22 |
+
"admitted_sets": [
|
| 23 |
+
"multihop",
|
| 24 |
+
"multilingual",
|
| 25 |
+
"order-ops",
|
| 26 |
+
"poetry",
|
| 27 |
+
"typo"
|
| 28 |
+
],
|
| 29 |
+
"K": [
|
| 30 |
+
1,
|
| 31 |
+
2,
|
| 32 |
+
5,
|
| 33 |
+
10,
|
| 34 |
+
20,
|
| 35 |
+
50,
|
| 36 |
+
100
|
| 37 |
+
],
|
| 38 |
+
"readout_convention": "STRIPPED (corrected); no read-context prefix",
|
| 39 |
+
"n_items": 449,
|
| 40 |
+
"n_pairs": 801,
|
| 41 |
+
"per_arm": {
|
| 42 |
+
"logit_I": {
|
| 43 |
+
"rank_min_auc": 0.1981085192666858,
|
| 44 |
+
"margin_mean": -4.698809347247424,
|
| 45 |
+
"z_mean": 3.398138034577217
|
| 46 |
+
},
|
| 47 |
+
"J|Wikipedia_en|s0": {
|
| 48 |
+
"rank_min_auc": 0.2411940539381176,
|
| 49 |
+
"margin_mean": -3.5185320037810324,
|
| 50 |
+
"z_mean": 3.414767909333594
|
| 51 |
+
},
|
| 52 |
+
"J|Wikipedia_en|s1": {
|
| 53 |
+
"rank_min_auc": 0.24376219820086684,
|
| 54 |
+
"margin_mean": -3.5576964909083792,
|
| 55 |
+
"z_mean": 3.4289134198996036
|
| 56 |
+
},
|
| 57 |
+
"J|Wikipedia_en|s2": {
|
| 58 |
+
"rank_min_auc": 0.2409968710017892,
|
| 59 |
+
"margin_mean": -3.550353089992781,
|
| 60 |
+
"z_mean": 3.418068973837049
|
| 61 |
+
},
|
| 62 |
+
"J|USPTO_Backgrounds|s0": {
|
| 63 |
+
"rank_min_auc": 0.2588173671008463,
|
| 64 |
+
"margin_mean": -3.5918301268675705,
|
| 65 |
+
"z_mean": 3.5110595907230495
|
| 66 |
+
},
|
| 67 |
+
"J|USPTO_Backgrounds|s1": {
|
| 68 |
+
"rank_min_auc": 0.260959847421275,
|
| 69 |
+
"margin_mean": -3.6230063218555544,
|
| 70 |
+
"z_mean": 3.5260322405353177
|
| 71 |
+
},
|
| 72 |
+
"J|USPTO_Backgrounds|s2": {
|
| 73 |
+
"rank_min_auc": 0.26360051696399417,
|
| 74 |
+
"margin_mean": -3.573313506298794,
|
| 75 |
+
"z_mean": 3.5303729015355394
|
| 76 |
+
},
|
| 77 |
+
"J|Pile-CC|s0": {
|
| 78 |
+
"rank_min_auc": 0.2619082412030062,
|
| 79 |
+
"margin_mean": -3.723244701992529,
|
| 80 |
+
"z_mean": 3.564116526229665
|
| 81 |
+
},
|
| 82 |
+
"J|Pile-CC|s1": {
|
| 83 |
+
"rank_min_auc": 0.2652928121057275,
|
| 84 |
+
"margin_mean": -3.721754557559919,
|
| 85 |
+
"z_mean": 3.57523909996598
|
| 86 |
+
},
|
| 87 |
+
"J|Pile-CC|s2": {
|
| 88 |
+
"rank_min_auc": 0.26001365602369086,
|
| 89 |
+
"margin_mean": -3.710869514133819,
|
| 90 |
+
"z_mean": 3.5523003176418593
|
| 91 |
+
},
|
| 92 |
+
"J|StackExchange|s0": {
|
| 93 |
+
"rank_min_auc": 0.26678756139973614,
|
| 94 |
+
"margin_mean": -3.8218988278155637,
|
| 95 |
+
"z_mean": 3.5992482651800937
|
| 96 |
+
},
|
| 97 |
+
"J|StackExchange|s1": {
|
| 98 |
+
"rank_min_auc": 0.263895495234026,
|
| 99 |
+
"margin_mean": -3.8108850501164118,
|
| 100 |
+
"z_mean": 3.5846069769433595
|
| 101 |
+
},
|
| 102 |
+
"J|StackExchange|s2": {
|
| 103 |
+
"rank_min_auc": 0.26310253043676124,
|
| 104 |
+
"margin_mean": -3.8273039062383045,
|
| 105 |
+
"z_mean": 3.591593915826658
|
| 106 |
+
},
|
| 107 |
+
"J|Github|s0": {
|
| 108 |
+
"rank_min_auc": 0.2361789172766393,
|
| 109 |
+
"margin_mean": -3.407463207637055,
|
| 110 |
+
"z_mean": 3.2020849497495396
|
| 111 |
+
},
|
| 112 |
+
"J|Github|s1": {
|
| 113 |
+
"rank_min_auc": 0.24343462540305366,
|
| 114 |
+
"margin_mean": -3.4403224028664545,
|
| 115 |
+
"z_mean": 3.2485827239444456
|
| 116 |
+
},
|
| 117 |
+
"J|Github|s2": {
|
| 118 |
+
"rank_min_auc": 0.2416357198287925,
|
| 119 |
+
"margin_mean": -3.410493464114918,
|
| 120 |
+
"z_mean": 3.232178426457889
|
| 121 |
+
},
|
| 122 |
+
"shuf|Pile-CC|s0|d0": {
|
| 123 |
+
"rank_min_auc": 0.27188047356388173,
|
| 124 |
+
"margin_mean": -3.078709140214322,
|
| 125 |
+
"z_mean": 3.5705154663621523
|
| 126 |
+
},
|
| 127 |
+
"shuf|Pile-CC|s0|d1": {
|
| 128 |
+
"rank_min_auc": 0.25473107820874097,
|
| 129 |
+
"margin_mean": -3.5329598556097404,
|
| 130 |
+
"z_mean": 3.4034280227130598
|
| 131 |
+
},
|
| 132 |
+
"shuf|Pile-CC|s0|d2": {
|
| 133 |
+
"rank_min_auc": 0.27573013496239485,
|
| 134 |
+
"margin_mean": -3.3755934340423375,
|
| 135 |
+
"z_mean": 3.461937350128892
|
| 136 |
+
}
|
| 137 |
+
},
|
| 138 |
+
"rank_space": {
|
| 139 |
+
"spread": 0.024178774854012652,
|
| 140 |
+
"pooled_seed_sd": 0.0025811352265854045,
|
| 141 |
+
"spread_over_sd": 9.36749636554256,
|
| 142 |
+
"per_corpus": {
|
| 143 |
+
"Wikipedia_en": {
|
| 144 |
+
"mean": 0.24198437438025788,
|
| 145 |
+
"seed_sd": 0.001542794033705807
|
| 146 |
+
},
|
| 147 |
+
"USPTO_Backgrounds": {
|
| 148 |
+
"mean": 0.2611259104953718,
|
| 149 |
+
"seed_sd": 0.002395895106586836
|
| 150 |
+
},
|
| 151 |
+
"Pile-CC": {
|
| 152 |
+
"mean": 0.2624049031108082,
|
| 153 |
+
"seed_sd": 0.0026743928325177486
|
| 154 |
+
},
|
| 155 |
+
"StackExchange": {
|
| 156 |
+
"mean": 0.2645951956901745,
|
| 157 |
+
"seed_sd": 0.0019396001767817256
|
| 158 |
+
},
|
| 159 |
+
"Github": {
|
| 160 |
+
"mean": 0.24041642083616183,
|
| 161 |
+
"seed_sd": 0.0037784047730492635
|
| 162 |
+
}
|
| 163 |
+
}
|
| 164 |
+
},
|
| 165 |
+
"margin_space": {
|
| 166 |
+
"spread": 0.40060290318395086,
|
| 167 |
+
"pooled_seed_sd": 0.01737544449501888,
|
| 168 |
+
"spread_over_sd": 23.055692376606196,
|
| 169 |
+
"per_corpus": {
|
| 170 |
+
"Wikipedia_en": {
|
| 171 |
+
"mean": -3.542193861560731,
|
| 172 |
+
"seed_sd": 0.020818117576808237
|
| 173 |
+
},
|
| 174 |
+
"USPTO_Backgrounds": {
|
| 175 |
+
"mean": -3.596049985007306,
|
| 176 |
+
"seed_sd": 0.025113728945124182
|
| 177 |
+
},
|
| 178 |
+
"Pile-CC": {
|
| 179 |
+
"mean": -3.718622924562089,
|
| 180 |
+
"seed_sd": 0.006755861348445229
|
| 181 |
+
},
|
| 182 |
+
"StackExchange": {
|
| 183 |
+
"mean": -3.8200292613900935,
|
| 184 |
+
"seed_sd": 0.008367566435486981
|
| 185 |
+
},
|
| 186 |
+
"Github": {
|
| 187 |
+
"mean": -3.4194263582061426,
|
| 188 |
+
"seed_sd": 0.018159821734111948
|
| 189 |
+
}
|
| 190 |
+
}
|
| 191 |
+
},
|
| 192 |
+
"z_space": {
|
| 193 |
+
"spread": 0.36420101926607895,
|
| 194 |
+
"pooled_seed_sd": 0.013407308018930473,
|
| 195 |
+
"spread_over_sd": 27.164365788556857,
|
| 196 |
+
"per_corpus": {
|
| 197 |
+
"Wikipedia_en": {
|
| 198 |
+
"mean": 3.420583434356749,
|
| 199 |
+
"seed_sd": 0.00740038857575187
|
| 200 |
+
},
|
| 201 |
+
"USPTO_Backgrounds": {
|
| 202 |
+
"mean": 3.5224882442646357,
|
| 203 |
+
"seed_sd": 0.010132666273489743
|
| 204 |
+
},
|
| 205 |
+
"Pile-CC": {
|
| 206 |
+
"mean": 3.5638853146125014,
|
| 207 |
+
"seed_sd": 0.011471138903233818
|
| 208 |
+
},
|
| 209 |
+
"StackExchange": {
|
| 210 |
+
"mean": 3.5918163859833703,
|
| 211 |
+
"seed_sd": 0.007323178956966246
|
| 212 |
+
},
|
| 213 |
+
"Github": {
|
| 214 |
+
"mean": 3.2276153667172913,
|
| 215 |
+
"seed_sd": 0.0235823426935772
|
| 216 |
+
}
|
| 217 |
+
}
|
| 218 |
+
},
|
| 219 |
+
"PRIMARY": 2.461243802710653,
|
| 220 |
+
"VERDICT": "ACCEPT \u2014 SCORE EFFECT",
|
| 221 |
+
"SECONDARY_derangement": {
|
| 222 |
+
"real_minus_shuf_rank_min_auc": -0.0055389877086662875,
|
| 223 |
+
"real_minus_shuf_margin": -0.394157225370396,
|
| 224 |
+
"real_minus_shuf_z": 0.08548957982829686,
|
| 225 |
+
"note": "if the derangement's advantage is present in rank space and absent in margin space, `min` rewards rank reshuffling rather than better scoring"
|
| 226 |
+
},
|
| 227 |
+
"controls": {
|
| 228 |
+
"C1_logit_reproduces_stored": {
|
| 229 |
+
"required": "|logit_I rank_min_auc - 0.19810852520167826| <= 1e-6",
|
| 230 |
+
"observed": 0.1981085192666858,
|
| 231 |
+
"abs_diff": 5.934992469702749e-09,
|
| 232 |
+
"fires": true
|
| 233 |
+
},
|
| 234 |
+
"C2_margin_rank_consistency": {
|
| 235 |
+
"required": "margin > 0 iff rank == 1 on 100% of triples",
|
| 236 |
+
"violations": 0,
|
| 237 |
+
"fires": true
|
| 238 |
+
},
|
| 239 |
+
"C4_seed_sd_nondegenerate": {
|
| 240 |
+
"required": "pooled seed SD > 0 in both spaces",
|
| 241 |
+
"rank": 0.0025811352265854045,
|
| 242 |
+
"margin": 0.01737544449501888,
|
| 243 |
+
"fires": true
|
| 244 |
+
}
|
| 245 |
+
},
|
| 246 |
+
"declared_bias": "margin is measured against the top competitor, which is itself operator-dependent; z and the raw logit are recorded so a competitor-suppression effect is separable from a target-score effect",
|
| 247 |
+
"provenance": {
|
| 248 |
+
"schema_version": 1,
|
| 249 |
+
"experiment": "CV3",
|
| 250 |
+
"utc": "2026-08-30T06:24:45+00:00",
|
| 251 |
+
"script": {
|
| 252 |
+
"path": "experiments/cv3_margins.py",
|
| 253 |
+
"sha256": "44938026f765b4f605189d29781ddb59a76d2a1bf91a230fabedb941a85e9521",
|
| 254 |
+
"n_lines": 247
|
| 255 |
+
},
|
| 256 |
+
"git": {
|
| 257 |
+
"commit": "c9c49697e295c0fec4225552836fc9733b279f75",
|
| 258 |
+
"commit_short": "c9c49697e295",
|
| 259 |
+
"git_repository_present": true,
|
| 260 |
+
"branch": "pythia",
|
| 261 |
+
"dirty": true,
|
| 262 |
+
"dirty_files": [
|
| 263 |
+
"CLAUDE.md",
|
| 264 |
+
"README.md",
|
| 265 |
+
"corpora/manifest_ood.json",
|
| 266 |
+
"docs/HANDOFF.md",
|
| 267 |
+
"docs/REPO_REVIEW_PROMPT.md",
|
| 268 |
+
"docs/context/AGGREGATION_POLICY.md",
|
| 269 |
+
"docs/context/CONTEXT.md",
|
| 270 |
+
"docs/context/IMPLEMENTATION_RULES.md",
|
| 271 |
+
"docs/context/STATE.md",
|
| 272 |
+
"docs/experiments/INDEX.md",
|
| 273 |
+
"docs/experiments/METRICS.md",
|
| 274 |
+
"docs/experiments/PROPOSED.md",
|
| 275 |
+
"docs/experiments/PROVENANCE.md",
|
| 276 |
+
"docs/experiments/descriptions/E64.md",
|
| 277 |
+
"docs/experiments/descriptions/EN1.md",
|
| 278 |
+
"docs/experiments/preregs/CV3_margins.md",
|
| 279 |
+
"docs/experiments/preregs/CV4_capability_ladder.md",
|
| 280 |
+
"docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 281 |
+
"docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 282 |
+
"docs/experiments/preregs/CV7_1b_rung.md",
|
| 283 |
+
"docs/experiments/preregs/DN1_design_nulls.md",
|
| 284 |
+
"docs/experiments/preregs/E33_logit_baseline.md",
|
| 285 |
+
"docs/experiments/preregs/E36_qladder.md",
|
| 286 |
+
"docs/experiments/preregs/E48_crossover.md",
|
| 287 |
+
"docs/experiments/preregs/E48b_containment.md",
|
| 288 |
+
"docs/experiments/preregs/E48c_exposure_vs_read.md",
|
| 289 |
+
"docs/experiments/preregs/E51_interaction.md",
|
| 290 |
+
"docs/experiments/preregs/E52_factorial.md",
|
| 291 |
+
"docs/experiments/preregs/E53_ladder_summary.md",
|
| 292 |
+
"docs/experiments/preregs/E55_matrix_robustness.md",
|
| 293 |
+
"docs/experiments/preregs/E56_predictor_registry.md",
|
| 294 |
+
"docs/experiments/preregs/E57_grid_variance_ci.md",
|
| 295 |
+
"docs/experiments/preregs/E58_algebra_audit.md",
|
| 296 |
+
"docs/experiments/preregs/E59_read_dose_response.md",
|
| 297 |
+
"docs/experiments/preregs/E60_fitter_determinism.md",
|
| 298 |
+
"docs/experiments/preregs/E61_randomized_network_null.md",
|
| 299 |
+
"docs/experiments/preregs/E62_ladder1b_corrected_band.md",
|
| 300 |
+
"docs/experiments/preregs/E65_training_axis_floor.md",
|
| 301 |
+
"docs/experiments/preregs/E66b_determinism_floor.md",
|
| 302 |
+
"docs/experiments/preregs/OS1_operator_space.md",
|
| 303 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md",
|
| 304 |
+
"docs/experiments/preregs/R1_grid_rstrip.md",
|
| 305 |
+
"docs/experiments/preregs/R2_e48_rstrip_arm.md",
|
| 306 |
+
"docs/experiments/preregs/R3_ladder410_cpu.md",
|
| 307 |
+
"docs/experiments/preregs/R4_corrections.md",
|
| 308 |
+
"docs/experiments/preregs/R4b_e36_rstrip.md",
|
| 309 |
+
"docs/experiments/preregs/R5_corpus_axis_uncertainty.md",
|
| 310 |
+
"docs/experiments/preregs/R6_within_source_resampling.md",
|
| 311 |
+
"docs/experiments/preregs/R8_ladder_rstrip_S2.md",
|
| 312 |
+
"docs/experiments/preregs/R9_permutation_calibrated_min.md",
|
| 313 |
+
"docs/experiments/preregs/superseded/PREREG_E48_CROSSOVER.md",
|
| 314 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 315 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 316 |
+
"docs/reproducibility/RELEASE.md",
|
| 317 |
+
"docs/reproducibility/REVIEWER_REPRO.md",
|
| 318 |
+
"docs/reproducibility/T1_SWEEP_PREDICTIONS.md",
|
| 319 |
+
"docs/reproducibility/T1_SWEEP_SPEC.md",
|
| 320 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 321 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 322 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 323 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 324 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 325 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md",
|
| 326 |
+
"docs/reproducibility/t1_module_audit.json"
|
| 327 |
+
],
|
| 328 |
+
"dirty_diff_sha256": "ec61dde1905616f407f51d5f6fe3ea37c9c6e3e1ade6998cd0644527d080c600",
|
| 329 |
+
"describe": "c9c4969-dirty"
|
| 330 |
+
},
|
| 331 |
+
"argv": [
|
| 332 |
+
"experiments/cv3_margins.py"
|
| 333 |
+
],
|
| 334 |
+
"cwd": ".",
|
| 335 |
+
"inputs": [
|
| 336 |
+
{
|
| 337 |
+
"path": "results/e48/lens_INSTREAM_Wikipedia_en_410m_n200_s0.pt",
|
| 338 |
+
"sha256": "50168cf38a9aab76190187b3669a4f26b0cf0ae69968e2decfee4e7ce0a26f8a",
|
| 339 |
+
"bytes": 27268444,
|
| 340 |
+
"exists": true
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"path": "results/e48/lens_INSTREAM_Wikipedia_en_410m_n200_s1.pt",
|
| 344 |
+
"sha256": "a6614fc65dd71659e2d1cb5799c37bffb291be1310fb06b69fb23921fcf4c9cc",
|
| 345 |
+
"bytes": 27268444,
|
| 346 |
+
"exists": true
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"path": "results/e48/lens_INSTREAM_Wikipedia_en_410m_n200_s2.pt",
|
| 350 |
+
"sha256": "c0702e058e3c41fb1e8b10020fa27e6ac96d34862f0648f9e55f13bd33776ed5",
|
| 351 |
+
"bytes": 27268444,
|
| 352 |
+
"exists": true
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"path": "results/e48/lens_INSTREAM_USPTO_Backgrounds_410m_n200_s0.pt",
|
| 356 |
+
"sha256": "75162b82ab15f9144b3776aabf62d0a1c732a474650e48b04c65bc77614aa942",
|
| 357 |
+
"bytes": 27268539,
|
| 358 |
+
"exists": true
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"path": "results/e48/lens_INSTREAM_USPTO_Backgrounds_410m_n200_s1.pt",
|
| 362 |
+
"sha256": "860d4c8da99be8789db610f1046fa97d8e5f6a0b7c0b2559a64825451dc33c5f",
|
| 363 |
+
"bytes": 27268539,
|
| 364 |
+
"exists": true
|
| 365 |
+
},
|
| 366 |
+
{
|
| 367 |
+
"path": "results/e48/lens_INSTREAM_USPTO_Backgrounds_410m_n200_s2.pt",
|
| 368 |
+
"sha256": "1a727ecb9d2a8dfcf76db138ef90c52cd2fe98bcef0be07ae759969b3d9c67e1",
|
| 369 |
+
"bytes": 27268539,
|
| 370 |
+
"exists": true
|
| 371 |
+
},
|
| 372 |
+
{
|
| 373 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s0.pt",
|
| 374 |
+
"sha256": "4f00cc3d7450c3646b78f002e794572ca27d6ab175c70626d2d5fb7fccf28e20",
|
| 375 |
+
"bytes": 27268349,
|
| 376 |
+
"exists": true
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s1.pt",
|
| 380 |
+
"sha256": "efc6a88531fdcd4423514fd0e4f65a02b55638653c351ad0af35cd236aad2fc2",
|
| 381 |
+
"bytes": 27268349,
|
| 382 |
+
"exists": true
|
| 383 |
+
},
|
| 384 |
+
{
|
| 385 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s2.pt",
|
| 386 |
+
"sha256": "058810065ae8bc499931d6aa3783dc1ead43d08e322395f7f142b90dfefae028",
|
| 387 |
+
"bytes": 27268349,
|
| 388 |
+
"exists": true
|
| 389 |
+
},
|
| 390 |
+
{
|
| 391 |
+
"path": "results/e48/lens_INSTREAM_StackExchange_410m_n200_s0.pt",
|
| 392 |
+
"sha256": "6ee98aa7e9242d0c4d6ff26b1ed918180f5b37822e30181c6d0fc9fe826c0b79",
|
| 393 |
+
"bytes": 27268463,
|
| 394 |
+
"exists": true
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"path": "results/e48/lens_INSTREAM_StackExchange_410m_n200_s1.pt",
|
| 398 |
+
"sha256": "9182d6655dc5808e47b92fe4611fc522226db741c4e6b900f8ba0e1eba7ca7b3",
|
| 399 |
+
"bytes": 27268463,
|
| 400 |
+
"exists": true
|
| 401 |
+
},
|
| 402 |
+
{
|
| 403 |
+
"path": "results/e48/lens_INSTREAM_StackExchange_410m_n200_s2.pt",
|
| 404 |
+
"sha256": "7e21bde15c9634ffff62ec1983ad595f0604915dc8a8f9169a165e33edaee8c0",
|
| 405 |
+
"bytes": 27268463,
|
| 406 |
+
"exists": true
|
| 407 |
+
},
|
| 408 |
+
{
|
| 409 |
+
"path": "results/e48/lens_INSTREAM_Github_410m_n200_s0.pt",
|
| 410 |
+
"sha256": "b4acd10402749791b713e16738ba1d165c04d7134df139418618fd05a4edcbdb",
|
| 411 |
+
"bytes": 27268330,
|
| 412 |
+
"exists": true
|
| 413 |
+
},
|
| 414 |
+
{
|
| 415 |
+
"path": "results/e48/lens_INSTREAM_Github_410m_n200_s1.pt",
|
| 416 |
+
"sha256": "0ec0403597cbdc971dc0348fae888a10aaa0602b8c9f9268ba008af83e07694b",
|
| 417 |
+
"bytes": 27268330,
|
| 418 |
+
"exists": true
|
| 419 |
+
},
|
| 420 |
+
{
|
| 421 |
+
"path": "results/e48/lens_INSTREAM_Github_410m_n200_s2.pt",
|
| 422 |
+
"sha256": "f5057978c3f1e94930b3c3892ef560c79cc2c68f90062d6b07aeefdec7828847",
|
| 423 |
+
"bytes": 27268330,
|
| 424 |
+
"exists": true
|
| 425 |
+
},
|
| 426 |
+
{
|
| 427 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s0.pt",
|
| 428 |
+
"sha256": "4f00cc3d7450c3646b78f002e794572ca27d6ab175c70626d2d5fb7fccf28e20",
|
| 429 |
+
"bytes": 27268349,
|
| 430 |
+
"exists": true
|
| 431 |
+
},
|
| 432 |
+
{
|
| 433 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s0.pt",
|
| 434 |
+
"sha256": "4f00cc3d7450c3646b78f002e794572ca27d6ab175c70626d2d5fb7fccf28e20",
|
| 435 |
+
"bytes": 27268349,
|
| 436 |
+
"exists": true
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s0.pt",
|
| 440 |
+
"sha256": "4f00cc3d7450c3646b78f002e794572ca27d6ab175c70626d2d5fb7fccf28e20",
|
| 441 |
+
"bytes": 27268349,
|
| 442 |
+
"exists": true
|
| 443 |
+
},
|
| 444 |
+
{
|
| 445 |
+
"path": "results/e48_crossover_410m_rstrip.json",
|
| 446 |
+
"sha256": "fe0b5ae29baf7f946884618723610d19120158474538e540eb1be8583430d97d",
|
| 447 |
+
"bytes": 111333,
|
| 448 |
+
"exists": true
|
| 449 |
+
}
|
| 450 |
+
],
|
| 451 |
+
"env": {
|
| 452 |
+
"python": "3.14.2",
|
| 453 |
+
"torch": "2.13.0",
|
| 454 |
+
"transformers": "5.14.1",
|
| 455 |
+
"numpy": "2.5.1",
|
| 456 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 457 |
+
"machine": "arm64",
|
| 458 |
+
"hostname": "<redacted for double-blind review>",
|
| 459 |
+
"tf32": {
|
| 460 |
+
"allow_tf32_matmul": false,
|
| 461 |
+
"allow_tf32_cudnn": true,
|
| 462 |
+
"NVIDIA_TF32_OVERRIDE": null
|
| 463 |
+
},
|
| 464 |
+
"threads": {
|
| 465 |
+
"torch_num_threads": 8,
|
| 466 |
+
"torch_num_interop_threads": 14,
|
| 467 |
+
"OMP_NUM_THREADS": null,
|
| 468 |
+
"os_cpu_count": 14
|
| 469 |
+
}
|
| 470 |
+
},
|
| 471 |
+
"payload_sha256": "38efe87c84be1b8bee5fa680a5a3d354912fedf8c7e29e686fd97fbecb9a8b5c"
|
| 472 |
+
}
|
| 473 |
+
}
|
results/cv4_phase1_capability.json
ADDED
|
@@ -0,0 +1,719 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV4 Phase 1 \u2014 capability calibration across the ladder, all six families",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV4_capability_ladder.md",
|
| 4 |
+
"status": "MEASUREMENT ONLY \u2014 Phase 1 gates which model later phases target; CV4's registered rule is adjudicated in Phase 3, not here",
|
| 5 |
+
"readout_convention": "STRIPPED (corrected)",
|
| 6 |
+
"n_items_total": 551,
|
| 7 |
+
"n_items_with_target": 255,
|
| 8 |
+
"measures": {
|
| 9 |
+
"answer_rank": "rank of the expected answer's first token; defined only where `target` is non-null (multihop, order-ops, multilingual)",
|
| 10 |
+
"prompt_surprisal": "mean per-token NLL of the prompt; defined for all six families. NOT a task-competence measure \u2014 it separates 'cannot do the task' from 'cannot parse the prompt'"
|
| 11 |
+
},
|
| 12 |
+
"note_on_undefined": "poetry / typo / association carry target: null. The annotated intermediate is the object of interest and there is no separate answer, so ANSWER COMPETENCE IS UNDEFINED for them, not zero.",
|
| 13 |
+
"interval": "95% Wilson score interval on every rate",
|
| 14 |
+
"by_model": {
|
| 15 |
+
"70m": {
|
| 16 |
+
"multihop": {
|
| 17 |
+
"n_items": 93,
|
| 18 |
+
"mean_prompt_surprisal": 4.742897197764407,
|
| 19 |
+
"median_prompt_surprisal": 4.761600017547607,
|
| 20 |
+
"n_scored_for_answer": 93,
|
| 21 |
+
"top1_count": 1,
|
| 22 |
+
"top1_rate": 0.010752688172043012,
|
| 23 |
+
"top1_wilson95": [
|
| 24 |
+
0.001900593097955601,
|
| 25 |
+
0.05842059118648415
|
| 26 |
+
],
|
| 27 |
+
"top10_count": 13,
|
| 28 |
+
"top10_rate": 0.13978494623655913,
|
| 29 |
+
"top10_wilson95": [
|
| 30 |
+
0.08354643741721157,
|
| 31 |
+
0.22460212683616715
|
| 32 |
+
],
|
| 33 |
+
"median_rank": 80
|
| 34 |
+
},
|
| 35 |
+
"multilingual": {
|
| 36 |
+
"n_items": 107,
|
| 37 |
+
"mean_prompt_surprisal": 5.743064506031643,
|
| 38 |
+
"median_prompt_surprisal": 5.629157066345215,
|
| 39 |
+
"n_scored_for_answer": 107,
|
| 40 |
+
"top1_count": 1,
|
| 41 |
+
"top1_rate": 0.009345794392523364,
|
| 42 |
+
"top1_wilson95": [
|
| 43 |
+
0.0016516291511190044,
|
| 44 |
+
0.05105060538898146
|
| 45 |
+
],
|
| 46 |
+
"top10_count": 9,
|
| 47 |
+
"top10_rate": 0.08411214953271028,
|
| 48 |
+
"top10_wilson95": [
|
| 49 |
+
0.044881436613206184,
|
| 50 |
+
0.15217093361602183
|
| 51 |
+
],
|
| 52 |
+
"median_rank": 3436
|
| 53 |
+
},
|
| 54 |
+
"order-ops": {
|
| 55 |
+
"n_items": 55,
|
| 56 |
+
"mean_prompt_surprisal": 4.685458686135032,
|
| 57 |
+
"median_prompt_surprisal": 4.4597907066345215,
|
| 58 |
+
"n_scored_for_answer": 55,
|
| 59 |
+
"top1_count": 2,
|
| 60 |
+
"top1_rate": 0.03636363636363636,
|
| 61 |
+
"top1_wilson95": [
|
| 62 |
+
0.010029330855724924,
|
| 63 |
+
0.1232369297490173
|
| 64 |
+
],
|
| 65 |
+
"top10_count": 21,
|
| 66 |
+
"top10_rate": 0.38181818181818183,
|
| 67 |
+
"top10_wilson95": [
|
| 68 |
+
0.26515766317415235,
|
| 69 |
+
0.513910207176076
|
| 70 |
+
],
|
| 71 |
+
"median_rank": 18
|
| 72 |
+
},
|
| 73 |
+
"poetry": {
|
| 74 |
+
"n_items": 98,
|
| 75 |
+
"mean_prompt_surprisal": 5.154343254712163,
|
| 76 |
+
"median_prompt_surprisal": 5.116718769073486,
|
| 77 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 78 |
+
},
|
| 79 |
+
"typo": {
|
| 80 |
+
"n_items": 96,
|
| 81 |
+
"mean_prompt_surprisal": 6.320249448219935,
|
| 82 |
+
"median_prompt_surprisal": 6.34007453918457,
|
| 83 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 84 |
+
},
|
| 85 |
+
"association": {
|
| 86 |
+
"n_items": 102,
|
| 87 |
+
"mean_prompt_surprisal": 4.577553527027953,
|
| 88 |
+
"median_prompt_surprisal": 4.60029411315918,
|
| 89 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 90 |
+
},
|
| 91 |
+
"_pooled_answer": {
|
| 92 |
+
"n": 255,
|
| 93 |
+
"top1_count": 4,
|
| 94 |
+
"top1_rate": 0.01568627450980392,
|
| 95 |
+
"top1_wilson95": [
|
| 96 |
+
0.006116456526536161,
|
| 97 |
+
0.03963198576403846
|
| 98 |
+
],
|
| 99 |
+
"top10_rate": 0.16862745098039217,
|
| 100 |
+
"top10_wilson95": [
|
| 101 |
+
0.12766685656458393,
|
| 102 |
+
0.2194241828973882
|
| 103 |
+
]
|
| 104 |
+
},
|
| 105 |
+
"_pooled_surprisal_all_six": 5.203927769981855
|
| 106 |
+
},
|
| 107 |
+
"160m": {
|
| 108 |
+
"multihop": {
|
| 109 |
+
"n_items": 93,
|
| 110 |
+
"mean_prompt_surprisal": 4.323448388807235,
|
| 111 |
+
"median_prompt_surprisal": 4.265374183654785,
|
| 112 |
+
"n_scored_for_answer": 93,
|
| 113 |
+
"top1_count": 3,
|
| 114 |
+
"top1_rate": 0.03225806451612903,
|
| 115 |
+
"top1_wilson95": [
|
| 116 |
+
0.011030412935116203,
|
| 117 |
+
0.09059533467748004
|
| 118 |
+
],
|
| 119 |
+
"top10_count": 24,
|
| 120 |
+
"top10_rate": 0.25806451612903225,
|
| 121 |
+
"top10_wilson95": [
|
| 122 |
+
0.17998399469266027,
|
| 123 |
+
0.35533966786557925
|
| 124 |
+
],
|
| 125 |
+
"median_rank": 42
|
| 126 |
+
},
|
| 127 |
+
"multilingual": {
|
| 128 |
+
"n_items": 107,
|
| 129 |
+
"mean_prompt_surprisal": 6.095417096236042,
|
| 130 |
+
"median_prompt_surprisal": 5.124428749084473,
|
| 131 |
+
"n_scored_for_answer": 107,
|
| 132 |
+
"top1_count": 1,
|
| 133 |
+
"top1_rate": 0.009345794392523364,
|
| 134 |
+
"top1_wilson95": [
|
| 135 |
+
0.0016516291511190044,
|
| 136 |
+
0.05105060538898146
|
| 137 |
+
],
|
| 138 |
+
"top10_count": 9,
|
| 139 |
+
"top10_rate": 0.08411214953271028,
|
| 140 |
+
"top10_wilson95": [
|
| 141 |
+
0.044881436613206184,
|
| 142 |
+
0.15217093361602183
|
| 143 |
+
],
|
| 144 |
+
"median_rank": 3788
|
| 145 |
+
},
|
| 146 |
+
"order-ops": {
|
| 147 |
+
"n_items": 55,
|
| 148 |
+
"mean_prompt_surprisal": 4.33481783433394,
|
| 149 |
+
"median_prompt_surprisal": 4.353231430053711,
|
| 150 |
+
"n_scored_for_answer": 55,
|
| 151 |
+
"top1_count": 0,
|
| 152 |
+
"top1_rate": 0.0,
|
| 153 |
+
"top1_wilson95": [
|
| 154 |
+
0.0,
|
| 155 |
+
0.06528714378942788
|
| 156 |
+
],
|
| 157 |
+
"top10_count": 24,
|
| 158 |
+
"top10_rate": 0.43636363636363634,
|
| 159 |
+
"top10_wilson95": [
|
| 160 |
+
0.31373238636001305,
|
| 161 |
+
0.5673041592131869
|
| 162 |
+
],
|
| 163 |
+
"median_rank": 13
|
| 164 |
+
},
|
| 165 |
+
"poetry": {
|
| 166 |
+
"n_items": 98,
|
| 167 |
+
"mean_prompt_surprisal": 4.700447340400851,
|
| 168 |
+
"median_prompt_surprisal": 4.699031829833984,
|
| 169 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 170 |
+
},
|
| 171 |
+
"typo": {
|
| 172 |
+
"n_items": 96,
|
| 173 |
+
"mean_prompt_surprisal": 5.784147493541241,
|
| 174 |
+
"median_prompt_surprisal": 5.752850532531738,
|
| 175 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 176 |
+
},
|
| 177 |
+
"association": {
|
| 178 |
+
"n_items": 102,
|
| 179 |
+
"mean_prompt_surprisal": 4.123158244525685,
|
| 180 |
+
"median_prompt_surprisal": 4.120728969573975,
|
| 181 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 182 |
+
},
|
| 183 |
+
"_pooled_answer": {
|
| 184 |
+
"n": 255,
|
| 185 |
+
"top1_count": 4,
|
| 186 |
+
"top1_rate": 0.01568627450980392,
|
| 187 |
+
"top1_wilson95": [
|
| 188 |
+
0.006116456526536161,
|
| 189 |
+
0.03963198576403846
|
| 190 |
+
],
|
| 191 |
+
"top10_rate": 0.2235294117647059,
|
| 192 |
+
"top10_wilson95": [
|
| 193 |
+
0.1767132210847063,
|
| 194 |
+
0.27855208402853665
|
| 195 |
+
]
|
| 196 |
+
},
|
| 197 |
+
"_pooled_surprisal_all_six": 4.893572732974166
|
| 198 |
+
},
|
| 199 |
+
"410m": {
|
| 200 |
+
"multihop": {
|
| 201 |
+
"n_items": 93,
|
| 202 |
+
"mean_prompt_surprisal": 3.8603221395964264,
|
| 203 |
+
"median_prompt_surprisal": 3.814849376678467,
|
| 204 |
+
"n_scored_for_answer": 93,
|
| 205 |
+
"top1_count": 6,
|
| 206 |
+
"top1_rate": 0.06451612903225806,
|
| 207 |
+
"top1_wilson95": [
|
| 208 |
+
0.029901024142029796,
|
| 209 |
+
0.13368156846280116
|
| 210 |
+
],
|
| 211 |
+
"top10_count": 31,
|
| 212 |
+
"top10_rate": 0.3333333333333333,
|
| 213 |
+
"top10_wilson95": [
|
| 214 |
+
0.24582243182201943,
|
| 215 |
+
0.4340672023847677
|
| 216 |
+
],
|
| 217 |
+
"median_rank": 31
|
| 218 |
+
},
|
| 219 |
+
"multilingual": {
|
| 220 |
+
"n_items": 107,
|
| 221 |
+
"mean_prompt_surprisal": 4.675777234763743,
|
| 222 |
+
"median_prompt_surprisal": 4.748653411865234,
|
| 223 |
+
"n_scored_for_answer": 107,
|
| 224 |
+
"top1_count": 1,
|
| 225 |
+
"top1_rate": 0.009345794392523364,
|
| 226 |
+
"top1_wilson95": [
|
| 227 |
+
0.0016516291511190044,
|
| 228 |
+
0.05105060538898146
|
| 229 |
+
],
|
| 230 |
+
"top10_count": 6,
|
| 231 |
+
"top10_rate": 0.056074766355140186,
|
| 232 |
+
"top10_wilson95": [
|
| 233 |
+
0.025950008789329404,
|
| 234 |
+
0.11697106055647577
|
| 235 |
+
],
|
| 236 |
+
"median_rank": 3044
|
| 237 |
+
},
|
| 238 |
+
"order-ops": {
|
| 239 |
+
"n_items": 55,
|
| 240 |
+
"mean_prompt_surprisal": 4.2336343418468125,
|
| 241 |
+
"median_prompt_surprisal": 4.275657653808594,
|
| 242 |
+
"n_scored_for_answer": 55,
|
| 243 |
+
"top1_count": 5,
|
| 244 |
+
"top1_rate": 0.09090909090909091,
|
| 245 |
+
"top1_wilson95": [
|
| 246 |
+
0.039457573336373924,
|
| 247 |
+
0.19577736249133978
|
| 248 |
+
],
|
| 249 |
+
"top10_count": 30,
|
| 250 |
+
"top10_rate": 0.5454545454545454,
|
| 251 |
+
"top10_wilson95": [
|
| 252 |
+
0.41522461348360484,
|
| 253 |
+
0.669749282535538
|
| 254 |
+
],
|
| 255 |
+
"median_rank": 9
|
| 256 |
+
},
|
| 257 |
+
"poetry": {
|
| 258 |
+
"n_items": 98,
|
| 259 |
+
"mean_prompt_surprisal": 4.148344592172272,
|
| 260 |
+
"median_prompt_surprisal": 4.126254558563232,
|
| 261 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 262 |
+
},
|
| 263 |
+
"typo": {
|
| 264 |
+
"n_items": 96,
|
| 265 |
+
"mean_prompt_surprisal": 5.49502004434665,
|
| 266 |
+
"median_prompt_surprisal": 5.293105602264404,
|
| 267 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 268 |
+
},
|
| 269 |
+
"association": {
|
| 270 |
+
"n_items": 102,
|
| 271 |
+
"mean_prompt_surprisal": 3.8331370657565547,
|
| 272 |
+
"median_prompt_surprisal": 3.8432066440582275,
|
| 273 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 274 |
+
},
|
| 275 |
+
"_pooled_answer": {
|
| 276 |
+
"n": 255,
|
| 277 |
+
"top1_count": 12,
|
| 278 |
+
"top1_rate": 0.047058823529411764,
|
| 279 |
+
"top1_wilson95": [
|
| 280 |
+
0.02712132140633029,
|
| 281 |
+
0.08044098697068483
|
| 282 |
+
],
|
| 283 |
+
"top10_rate": 0.2627450980392157,
|
| 284 |
+
"top10_wilson95": [
|
| 285 |
+
0.21253225697118083,
|
| 286 |
+
0.3200003807501127
|
| 287 |
+
]
|
| 288 |
+
},
|
| 289 |
+
"_pooled_surprisal_all_six": 4.374372569747076
|
| 290 |
+
},
|
| 291 |
+
"1b": {
|
| 292 |
+
"multihop": {
|
| 293 |
+
"n_items": 93,
|
| 294 |
+
"mean_prompt_surprisal": 3.7122991469598587,
|
| 295 |
+
"median_prompt_surprisal": 3.713099241256714,
|
| 296 |
+
"n_scored_for_answer": 93,
|
| 297 |
+
"top1_count": 7,
|
| 298 |
+
"top1_rate": 0.07526881720430108,
|
| 299 |
+
"top1_wilson95": [
|
| 300 |
+
0.036936214085016475,
|
| 301 |
+
0.14729866018389276
|
| 302 |
+
],
|
| 303 |
+
"top10_count": 31,
|
| 304 |
+
"top10_rate": 0.3333333333333333,
|
| 305 |
+
"top10_wilson95": [
|
| 306 |
+
0.24582243182201943,
|
| 307 |
+
0.4340672023847677
|
| 308 |
+
],
|
| 309 |
+
"median_rank": 18
|
| 310 |
+
},
|
| 311 |
+
"multilingual": {
|
| 312 |
+
"n_items": 107,
|
| 313 |
+
"mean_prompt_surprisal": 4.4941605663745205,
|
| 314 |
+
"median_prompt_surprisal": 4.559407711029053,
|
| 315 |
+
"n_scored_for_answer": 107,
|
| 316 |
+
"top1_count": 0,
|
| 317 |
+
"top1_rate": 0.0,
|
| 318 |
+
"top1_wilson95": [
|
| 319 |
+
0.0,
|
| 320 |
+
0.03465846757895952
|
| 321 |
+
],
|
| 322 |
+
"top10_count": 7,
|
| 323 |
+
"top10_rate": 0.06542056074766354,
|
| 324 |
+
"top10_wilson95": [
|
| 325 |
+
0.032047896397421095,
|
| 326 |
+
0.12891693990952502
|
| 327 |
+
],
|
| 328 |
+
"median_rank": 1649
|
| 329 |
+
},
|
| 330 |
+
"order-ops": {
|
| 331 |
+
"n_items": 55,
|
| 332 |
+
"mean_prompt_surprisal": 4.145381216569381,
|
| 333 |
+
"median_prompt_surprisal": 4.083983898162842,
|
| 334 |
+
"n_scored_for_answer": 55,
|
| 335 |
+
"top1_count": 4,
|
| 336 |
+
"top1_rate": 0.07272727272727272,
|
| 337 |
+
"top1_wilson95": [
|
| 338 |
+
0.028643596503688543,
|
| 339 |
+
0.17260178091636796
|
| 340 |
+
],
|
| 341 |
+
"top10_count": 37,
|
| 342 |
+
"top10_rate": 0.6727272727272727,
|
| 343 |
+
"top10_wilson95": [
|
| 344 |
+
0.5410297617609947,
|
| 345 |
+
0.7818710431117484
|
| 346 |
+
],
|
| 347 |
+
"median_rank": 7
|
| 348 |
+
},
|
| 349 |
+
"poetry": {
|
| 350 |
+
"n_items": 98,
|
| 351 |
+
"mean_prompt_surprisal": 3.935050759996687,
|
| 352 |
+
"median_prompt_surprisal": 3.910858154296875,
|
| 353 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 354 |
+
},
|
| 355 |
+
"typo": {
|
| 356 |
+
"n_items": 96,
|
| 357 |
+
"mean_prompt_surprisal": 5.290679027636846,
|
| 358 |
+
"median_prompt_surprisal": 5.2639923095703125,
|
| 359 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 360 |
+
},
|
| 361 |
+
"association": {
|
| 362 |
+
"n_items": 102,
|
| 363 |
+
"mean_prompt_surprisal": 3.6671034354789582,
|
| 364 |
+
"median_prompt_surprisal": 3.6480932235717773,
|
| 365 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 366 |
+
},
|
| 367 |
+
"_pooled_answer": {
|
| 368 |
+
"n": 255,
|
| 369 |
+
"top1_count": 11,
|
| 370 |
+
"top1_rate": 0.043137254901960784,
|
| 371 |
+
"top1_wilson95": [
|
| 372 |
+
0.024254970635041635,
|
| 373 |
+
0.07558060448116843
|
| 374 |
+
],
|
| 375 |
+
"top10_rate": 0.29411764705882354,
|
| 376 |
+
"top10_wilson95": [
|
| 377 |
+
0.24157991671293233,
|
| 378 |
+
0.3527665870948018
|
| 379 |
+
]
|
| 380 |
+
},
|
| 381 |
+
"_pooled_surprisal_all_six": 4.207445692169375
|
| 382 |
+
},
|
| 383 |
+
"1.4b": {
|
| 384 |
+
"multihop": {
|
| 385 |
+
"n_items": 93,
|
| 386 |
+
"mean_prompt_surprisal": 3.5891012683991463,
|
| 387 |
+
"median_prompt_surprisal": 3.6213531494140625,
|
| 388 |
+
"n_scored_for_answer": 93,
|
| 389 |
+
"top1_count": 5,
|
| 390 |
+
"top1_rate": 0.053763440860215055,
|
| 391 |
+
"top1_wilson95": [
|
| 392 |
+
0.023180332422992958,
|
| 393 |
+
0.11974997851775977
|
| 394 |
+
],
|
| 395 |
+
"top10_count": 33,
|
| 396 |
+
"top10_rate": 0.3548387096774194,
|
| 397 |
+
"top10_wilson95": [
|
| 398 |
+
0.26512726316736557,
|
| 399 |
+
0.45606693436757817
|
| 400 |
+
],
|
| 401 |
+
"median_rank": 24
|
| 402 |
+
},
|
| 403 |
+
"multilingual": {
|
| 404 |
+
"n_items": 107,
|
| 405 |
+
"mean_prompt_surprisal": 4.305595877014588,
|
| 406 |
+
"median_prompt_surprisal": 4.307002067565918,
|
| 407 |
+
"n_scored_for_answer": 107,
|
| 408 |
+
"top1_count": 0,
|
| 409 |
+
"top1_rate": 0.0,
|
| 410 |
+
"top1_wilson95": [
|
| 411 |
+
0.0,
|
| 412 |
+
0.03465846757895952
|
| 413 |
+
],
|
| 414 |
+
"top10_count": 8,
|
| 415 |
+
"top10_rate": 0.07476635514018691,
|
| 416 |
+
"top10_wilson95": [
|
| 417 |
+
0.03836964356083345,
|
| 418 |
+
0.1406389597072536
|
| 419 |
+
],
|
| 420 |
+
"median_rank": 1345
|
| 421 |
+
},
|
| 422 |
+
"order-ops": {
|
| 423 |
+
"n_items": 55,
|
| 424 |
+
"mean_prompt_surprisal": 4.166462430087003,
|
| 425 |
+
"median_prompt_surprisal": 4.244797229766846,
|
| 426 |
+
"n_scored_for_answer": 55,
|
| 427 |
+
"top1_count": 3,
|
| 428 |
+
"top1_rate": 0.05454545454545454,
|
| 429 |
+
"top1_wilson95": [
|
| 430 |
+
0.01872287596750008,
|
| 431 |
+
0.14853294304489928
|
| 432 |
+
],
|
| 433 |
+
"top10_count": 29,
|
| 434 |
+
"top10_rate": 0.5272727272727272,
|
| 435 |
+
"top10_wilson95": [
|
| 436 |
+
0.39791319071629383,
|
| 437 |
+
0.6530711468951917
|
| 438 |
+
],
|
| 439 |
+
"median_rank": 7
|
| 440 |
+
},
|
| 441 |
+
"poetry": {
|
| 442 |
+
"n_items": 98,
|
| 443 |
+
"mean_prompt_surprisal": 3.6910850977411074,
|
| 444 |
+
"median_prompt_surprisal": 3.6510846614837646,
|
| 445 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 446 |
+
},
|
| 447 |
+
"typo": {
|
| 448 |
+
"n_items": 96,
|
| 449 |
+
"mean_prompt_surprisal": 5.277833836774032,
|
| 450 |
+
"median_prompt_surprisal": 5.16269588470459,
|
| 451 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 452 |
+
},
|
| 453 |
+
"association": {
|
| 454 |
+
"n_items": 102,
|
| 455 |
+
"mean_prompt_surprisal": 3.6086690916734585,
|
| 456 |
+
"median_prompt_surprisal": 3.6544148921966553,
|
| 457 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 458 |
+
},
|
| 459 |
+
"_pooled_answer": {
|
| 460 |
+
"n": 255,
|
| 461 |
+
"top1_count": 8,
|
| 462 |
+
"top1_rate": 0.03137254901960784,
|
| 463 |
+
"top1_wilson95": [
|
| 464 |
+
0.01598082278729794,
|
| 465 |
+
0.060674552546496924
|
| 466 |
+
],
|
| 467 |
+
"top10_rate": 0.27450980392156865,
|
| 468 |
+
"top10_wilson95": [
|
| 469 |
+
0.22338671792303816,
|
| 470 |
+
0.33232611958067076
|
| 471 |
+
]
|
| 472 |
+
},
|
| 473 |
+
"_pooled_surprisal_all_six": 4.106457933614889
|
| 474 |
+
},
|
| 475 |
+
"2.8b": {
|
| 476 |
+
"multihop": {
|
| 477 |
+
"n_items": 93,
|
| 478 |
+
"mean_prompt_surprisal": 3.5109470018776516,
|
| 479 |
+
"median_prompt_surprisal": 3.531287670135498,
|
| 480 |
+
"n_scored_for_answer": 93,
|
| 481 |
+
"top1_count": 6,
|
| 482 |
+
"top1_rate": 0.06451612903225806,
|
| 483 |
+
"top1_wilson95": [
|
| 484 |
+
0.029901024142029796,
|
| 485 |
+
0.13368156846280116
|
| 486 |
+
],
|
| 487 |
+
"top10_count": 40,
|
| 488 |
+
"top10_rate": 0.43010752688172044,
|
| 489 |
+
"top10_wilson95": [
|
| 490 |
+
0.33423358533237524,
|
| 491 |
+
0.5315265838511162
|
| 492 |
+
],
|
| 493 |
+
"median_rank": 14
|
| 494 |
+
},
|
| 495 |
+
"multilingual": {
|
| 496 |
+
"n_items": 107,
|
| 497 |
+
"mean_prompt_surprisal": 4.153321266174316,
|
| 498 |
+
"median_prompt_surprisal": 4.131369590759277,
|
| 499 |
+
"n_scored_for_answer": 107,
|
| 500 |
+
"top1_count": 1,
|
| 501 |
+
"top1_rate": 0.009345794392523364,
|
| 502 |
+
"top1_wilson95": [
|
| 503 |
+
0.0016516291511190044,
|
| 504 |
+
0.05105060538898146
|
| 505 |
+
],
|
| 506 |
+
"top10_count": 10,
|
| 507 |
+
"top10_rate": 0.09345794392523364,
|
| 508 |
+
"top10_wilson95": [
|
| 509 |
+
0.05155771001191554,
|
| 510 |
+
0.1635384271784534
|
| 511 |
+
],
|
| 512 |
+
"median_rank": 578
|
| 513 |
+
},
|
| 514 |
+
"order-ops": {
|
| 515 |
+
"n_items": 55,
|
| 516 |
+
"mean_prompt_surprisal": 4.071762145649303,
|
| 517 |
+
"median_prompt_surprisal": 4.097884178161621,
|
| 518 |
+
"n_scored_for_answer": 55,
|
| 519 |
+
"top1_count": 8,
|
| 520 |
+
"top1_rate": 0.14545454545454545,
|
| 521 |
+
"top1_wilson95": [
|
| 522 |
+
0.07559199016771778,
|
| 523 |
+
0.2616116208829674
|
| 524 |
+
],
|
| 525 |
+
"top10_count": 34,
|
| 526 |
+
"top10_rate": 0.6181818181818182,
|
| 527 |
+
"top10_wilson95": [
|
| 528 |
+
0.4860897928239239,
|
| 529 |
+
0.7348423368258477
|
| 530 |
+
],
|
| 531 |
+
"median_rank": 7
|
| 532 |
+
},
|
| 533 |
+
"poetry": {
|
| 534 |
+
"n_items": 98,
|
| 535 |
+
"mean_prompt_surprisal": 3.6883874936979644,
|
| 536 |
+
"median_prompt_surprisal": 3.6747243404388428,
|
| 537 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 538 |
+
},
|
| 539 |
+
"typo": {
|
| 540 |
+
"n_items": 96,
|
| 541 |
+
"mean_prompt_surprisal": 5.161082645257314,
|
| 542 |
+
"median_prompt_surprisal": 5.04867696762085,
|
| 543 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 544 |
+
},
|
| 545 |
+
"association": {
|
| 546 |
+
"n_items": 102,
|
| 547 |
+
"mean_prompt_surprisal": 3.4943248594508454,
|
| 548 |
+
"median_prompt_surprisal": 3.5100271701812744,
|
| 549 |
+
"answer_competence": "UNDEFINED \u2014 target is null for this family"
|
| 550 |
+
},
|
| 551 |
+
"_pooled_answer": {
|
| 552 |
+
"n": 255,
|
| 553 |
+
"top1_count": 15,
|
| 554 |
+
"top1_rate": 0.058823529411764705,
|
| 555 |
+
"top1_wilson95": [
|
| 556 |
+
0.03596805781169063,
|
| 557 |
+
0.09477445034773968
|
| 558 |
+
],
|
| 559 |
+
"top10_rate": 0.32941176470588235,
|
| 560 |
+
"top10_wilson95": [
|
| 561 |
+
0.2746295550144033,
|
| 562 |
+
0.3892575481405764
|
| 563 |
+
]
|
| 564 |
+
},
|
| 565 |
+
"_pooled_surprisal_all_six": 4.0133042353512325
|
| 566 |
+
}
|
| 567 |
+
},
|
| 568 |
+
"VERDICT": "Answer competence, 255 items with a target, 95% Wilson: 70m: top1 1.6% [0.6,4.0] | 160m: top1 1.6% [0.6,4.0] | 410m: top1 4.7% [2.7,8.0] | 1b: top1 4.3% [2.4,7.6] | 1.4b: top1 3.1% [1.6,6.1] | 2.8b: top1 5.9% [3.6,9.5]. Prompt surprisal covers all 551 items in all six families.",
|
| 569 |
+
"provenance": {
|
| 570 |
+
"schema_version": 1,
|
| 571 |
+
"experiment": "CV4P1",
|
| 572 |
+
"utc": "2026-08-30T06:30:11+00:00",
|
| 573 |
+
"script": {
|
| 574 |
+
"path": "experiments/cv4_phase1_capability.py",
|
| 575 |
+
"sha256": "3aaadc49c285d30fe1aafb92714633b4d5e694caac40f82b0afbfa964a424cfb",
|
| 576 |
+
"n_lines": 173
|
| 577 |
+
},
|
| 578 |
+
"git": {
|
| 579 |
+
"commit": "c9c49697e295c0fec4225552836fc9733b279f75",
|
| 580 |
+
"commit_short": "c9c49697e295",
|
| 581 |
+
"git_repository_present": true,
|
| 582 |
+
"branch": "pythia",
|
| 583 |
+
"dirty": true,
|
| 584 |
+
"dirty_files": [
|
| 585 |
+
"CLAUDE.md",
|
| 586 |
+
"README.md",
|
| 587 |
+
"corpora/manifest_ood.json",
|
| 588 |
+
"docs/HANDOFF.md",
|
| 589 |
+
"docs/REPO_REVIEW_PROMPT.md",
|
| 590 |
+
"docs/context/AGGREGATION_POLICY.md",
|
| 591 |
+
"docs/context/CONTEXT.md",
|
| 592 |
+
"docs/context/IMPLEMENTATION_RULES.md",
|
| 593 |
+
"docs/context/STATE.md",
|
| 594 |
+
"docs/experiments/INDEX.md",
|
| 595 |
+
"docs/experiments/METRICS.md",
|
| 596 |
+
"docs/experiments/PROPOSED.md",
|
| 597 |
+
"docs/experiments/PROVENANCE.md",
|
| 598 |
+
"docs/experiments/descriptions/E64.md",
|
| 599 |
+
"docs/experiments/descriptions/EN1.md",
|
| 600 |
+
"docs/experiments/preregs/CV3_margins.md",
|
| 601 |
+
"docs/experiments/preregs/CV4_capability_ladder.md",
|
| 602 |
+
"docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 603 |
+
"docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 604 |
+
"docs/experiments/preregs/CV7_1b_rung.md",
|
| 605 |
+
"docs/experiments/preregs/DN1_design_nulls.md",
|
| 606 |
+
"docs/experiments/preregs/E33_logit_baseline.md",
|
| 607 |
+
"docs/experiments/preregs/E36_qladder.md",
|
| 608 |
+
"docs/experiments/preregs/E48_crossover.md",
|
| 609 |
+
"docs/experiments/preregs/E48b_containment.md",
|
| 610 |
+
"docs/experiments/preregs/E48c_exposure_vs_read.md",
|
| 611 |
+
"docs/experiments/preregs/E51_interaction.md",
|
| 612 |
+
"docs/experiments/preregs/E52_factorial.md",
|
| 613 |
+
"docs/experiments/preregs/E53_ladder_summary.md",
|
| 614 |
+
"docs/experiments/preregs/E55_matrix_robustness.md",
|
| 615 |
+
"docs/experiments/preregs/E56_predictor_registry.md",
|
| 616 |
+
"docs/experiments/preregs/E57_grid_variance_ci.md",
|
| 617 |
+
"docs/experiments/preregs/E58_algebra_audit.md",
|
| 618 |
+
"docs/experiments/preregs/E59_read_dose_response.md",
|
| 619 |
+
"docs/experiments/preregs/E60_fitter_determinism.md",
|
| 620 |
+
"docs/experiments/preregs/E61_randomized_network_null.md",
|
| 621 |
+
"docs/experiments/preregs/E62_ladder1b_corrected_band.md",
|
| 622 |
+
"docs/experiments/preregs/E65_training_axis_floor.md",
|
| 623 |
+
"docs/experiments/preregs/E66b_determinism_floor.md",
|
| 624 |
+
"docs/experiments/preregs/OS1_operator_space.md",
|
| 625 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md",
|
| 626 |
+
"docs/experiments/preregs/R1_grid_rstrip.md",
|
| 627 |
+
"docs/experiments/preregs/R2_e48_rstrip_arm.md",
|
| 628 |
+
"docs/experiments/preregs/R3_ladder410_cpu.md",
|
| 629 |
+
"docs/experiments/preregs/R4_corrections.md",
|
| 630 |
+
"docs/experiments/preregs/R4b_e36_rstrip.md",
|
| 631 |
+
"docs/experiments/preregs/R5_corpus_axis_uncertainty.md",
|
| 632 |
+
"docs/experiments/preregs/R6_within_source_resampling.md",
|
| 633 |
+
"docs/experiments/preregs/R8_ladder_rstrip_S2.md",
|
| 634 |
+
"docs/experiments/preregs/R9_permutation_calibrated_min.md",
|
| 635 |
+
"docs/experiments/preregs/superseded/PREREG_E48_CROSSOVER.md",
|
| 636 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 637 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 638 |
+
"docs/reproducibility/RELEASE.md",
|
| 639 |
+
"docs/reproducibility/REVIEWER_REPRO.md",
|
| 640 |
+
"docs/reproducibility/T1_FINDING_E33_BAND.md",
|
| 641 |
+
"docs/reproducibility/T1_SWEEP_PREDICTIONS.md",
|
| 642 |
+
"docs/reproducibility/T1_SWEEP_SPEC.md",
|
| 643 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 644 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 645 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 646 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 647 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 648 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md"
|
| 649 |
+
],
|
| 650 |
+
"dirty_diff_sha256": "ec61dde1905616f407f51d5f6fe3ea37c9c6e3e1ade6998cd0644527d080c600",
|
| 651 |
+
"describe": "c9c4969-dirty"
|
| 652 |
+
},
|
| 653 |
+
"argv": [
|
| 654 |
+
"experiments/cv4_phase1_capability.py",
|
| 655 |
+
"--models",
|
| 656 |
+
"70m,160m,410m,1b,1.4b,2.8b"
|
| 657 |
+
],
|
| 658 |
+
"cwd": ".",
|
| 659 |
+
"inputs": [
|
| 660 |
+
{
|
| 661 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-multihop.json",
|
| 662 |
+
"sha256": "50b7e4c9255291c0ca2a8e94615be9f44531fa57bb1a844e4f9616056d987416",
|
| 663 |
+
"bytes": 21869,
|
| 664 |
+
"exists": true
|
| 665 |
+
},
|
| 666 |
+
{
|
| 667 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-multilingual.json",
|
| 668 |
+
"sha256": "fa70b9bd89416a6d8d985a80dc628b109ae6fd3b25b9275c0fc5065d7ff4a0ef",
|
| 669 |
+
"bytes": 24284,
|
| 670 |
+
"exists": true
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-order-ops.json",
|
| 674 |
+
"sha256": "b203206d16ff628152cc86f3838604e06cb54776f3e14fa1c34f150db8bc7560",
|
| 675 |
+
"bytes": 9589,
|
| 676 |
+
"exists": true
|
| 677 |
+
},
|
| 678 |
+
{
|
| 679 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-poetry.json",
|
| 680 |
+
"sha256": "6aeb3415c5a5c3f3827c9efe63f006de02f5ef39a816bbac68e15e733aba60cc",
|
| 681 |
+
"bytes": 21533,
|
| 682 |
+
"exists": true
|
| 683 |
+
},
|
| 684 |
+
{
|
| 685 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-typo.json",
|
| 686 |
+
"sha256": "9d05e16b7234a57d0773d120a4e1c4e94fd3bc2235a8125d4200a70e60ab17aa",
|
| 687 |
+
"bytes": 14646,
|
| 688 |
+
"exists": true
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"path": "jacobian-lens/data/evaluations/lens-eval-association.json",
|
| 692 |
+
"sha256": "d1a98cd4911b594282e74168091c77d849dae18ffe2acb5761074853f327d71c",
|
| 693 |
+
"bytes": 24228,
|
| 694 |
+
"exists": true
|
| 695 |
+
}
|
| 696 |
+
],
|
| 697 |
+
"env": {
|
| 698 |
+
"python": "3.14.2",
|
| 699 |
+
"torch": "2.13.0",
|
| 700 |
+
"transformers": "5.14.1",
|
| 701 |
+
"numpy": "2.5.1",
|
| 702 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 703 |
+
"machine": "arm64",
|
| 704 |
+
"hostname": "<redacted for double-blind review>",
|
| 705 |
+
"tf32": {
|
| 706 |
+
"allow_tf32_matmul": false,
|
| 707 |
+
"allow_tf32_cudnn": true,
|
| 708 |
+
"NVIDIA_TF32_OVERRIDE": null
|
| 709 |
+
},
|
| 710 |
+
"threads": {
|
| 711 |
+
"torch_num_threads": 10,
|
| 712 |
+
"torch_num_interop_threads": 14,
|
| 713 |
+
"OMP_NUM_THREADS": null,
|
| 714 |
+
"os_cpu_count": 14
|
| 715 |
+
}
|
| 716 |
+
},
|
| 717 |
+
"payload_sha256": "43b21658bcc6a1acf619b614a5eb74db9252372b3ff7f151fbbde0ce7fbd9bad"
|
| 718 |
+
}
|
| 719 |
+
}
|
results/cv5_metric_sensitivity_410m.json
ADDED
|
@@ -0,0 +1,296 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV5 \u2014 metric sensitivity: is rank blind where z is sighted?",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 4 |
+
"status": "PRE-REGISTERED",
|
| 5 |
+
"decision_rule_verbatim": "at r=2.535e-03: ACCEPT (RANK WAS BLIND) if S_z >= 1.0 and S_rank < 0.5 | REJECT (FUNCTIONALLY IDENTICAL) if S_z < 0.5 | UNCLEAR otherwise",
|
| 6 |
+
"model": "EleutherAI/pythia-410m-deduped",
|
| 7 |
+
"band": [
|
| 8 |
+
9,
|
| 9 |
+
10,
|
| 10 |
+
11,
|
| 11 |
+
12,
|
| 12 |
+
13,
|
| 13 |
+
14,
|
| 14 |
+
15,
|
| 15 |
+
16,
|
| 16 |
+
17,
|
| 17 |
+
18,
|
| 18 |
+
19,
|
| 19 |
+
20,
|
| 20 |
+
21
|
| 21 |
+
],
|
| 22 |
+
"admitted_sets": [
|
| 23 |
+
"multihop",
|
| 24 |
+
"multilingual",
|
| 25 |
+
"order-ops",
|
| 26 |
+
"poetry",
|
| 27 |
+
"typo"
|
| 28 |
+
],
|
| 29 |
+
"K": [
|
| 30 |
+
1,
|
| 31 |
+
2,
|
| 32 |
+
5,
|
| 33 |
+
10,
|
| 34 |
+
20,
|
| 35 |
+
50,
|
| 36 |
+
100
|
| 37 |
+
],
|
| 38 |
+
"base_operator": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s0.pt",
|
| 39 |
+
"readout_convention": "STRIPPED (corrected), no prefix; flat-mean-7 k-summary",
|
| 40 |
+
"denominators_from_cv3": {
|
| 41 |
+
"pooled_seed_sd_rank": 0.002581,
|
| 42 |
+
"pooled_seed_sd_z": 0.013407
|
| 43 |
+
},
|
| 44 |
+
"unperturbed": {
|
| 45 |
+
"rank": 0.2619082412030062,
|
| 46 |
+
"z": 3.564116526229665,
|
| 47 |
+
"logit_I_rank": 0.1981085192666858
|
| 48 |
+
},
|
| 49 |
+
"by_magnitude": {
|
| 50 |
+
"0.0": {
|
| 51 |
+
"r": 0.0,
|
| 52 |
+
"n_seeds": 1,
|
| 53 |
+
"rank": [
|
| 54 |
+
0.2619082412030062
|
| 55 |
+
],
|
| 56 |
+
"z": [
|
| 57 |
+
3.564116526229665
|
| 58 |
+
],
|
| 59 |
+
"S_rank": 0.0,
|
| 60 |
+
"S_z": 0.0
|
| 61 |
+
},
|
| 62 |
+
"1.06e-07": {
|
| 63 |
+
"r": 1.06e-07,
|
| 64 |
+
"n_seeds": 3,
|
| 65 |
+
"rank": [
|
| 66 |
+
0.2619082412030062,
|
| 67 |
+
0.2619082412030062,
|
| 68 |
+
0.2619082412030062
|
| 69 |
+
],
|
| 70 |
+
"z": [
|
| 71 |
+
3.5641165554510494,
|
| 72 |
+
3.5641163831333875,
|
| 73 |
+
3.5641164395128424
|
| 74 |
+
],
|
| 75 |
+
"S_rank": 0.0,
|
| 76 |
+
"S_z": 4.9872383888545455e-06
|
| 77 |
+
},
|
| 78 |
+
"0.002535": {
|
| 79 |
+
"r": 0.002535,
|
| 80 |
+
"n_seeds": 3,
|
| 81 |
+
"rank": [
|
| 82 |
+
0.2617184725746441,
|
| 83 |
+
0.2626325979167178,
|
| 84 |
+
0.2620876530516675
|
| 85 |
+
],
|
| 86 |
+
"z": [
|
| 87 |
+
3.562412863826774,
|
| 88 |
+
3.5641458966086614,
|
| 89 |
+
3.5612677342495367
|
| 90 |
+
],
|
| 91 |
+
"S_rank": 0.09221231228343889,
|
| 92 |
+
"S_z": 0.11245578190554997
|
| 93 |
+
},
|
| 94 |
+
"0.01": {
|
| 95 |
+
"r": 0.01,
|
| 96 |
+
"n_seeds": 3,
|
| 97 |
+
"rank": [
|
| 98 |
+
0.2663647765893257,
|
| 99 |
+
0.2655696648942096,
|
| 100 |
+
0.26799035067682875
|
| 101 |
+
],
|
| 102 |
+
"z": [
|
| 103 |
+
3.573528114914198,
|
| 104 |
+
3.562376512847641,
|
| 105 |
+
3.5774573911741827
|
| 106 |
+
],
|
| 107 |
+
"S_rank": 1.8339233567538897,
|
| 108 |
+
"S_z": 0.5224246102042763
|
| 109 |
+
},
|
| 110 |
+
"0.05": {
|
| 111 |
+
"r": 0.05,
|
| 112 |
+
"n_seeds": 3,
|
| 113 |
+
"rank": [
|
| 114 |
+
0.28435305275900635,
|
| 115 |
+
0.27584753020009956,
|
| 116 |
+
0.2805677163156452
|
| 117 |
+
],
|
| 118 |
+
"z": [
|
| 119 |
+
3.495605627184469,
|
| 120 |
+
3.473200516113597,
|
| 121 |
+
3.4979920391660375
|
| 122 |
+
],
|
| 123 |
+
"S_rank": 7.108817727719558,
|
| 124 |
+
"S_z": 5.607801800673574
|
| 125 |
+
}
|
| 126 |
+
},
|
| 127 |
+
"PRIMARY_at_e66_magnitude": {
|
| 128 |
+
"r": 0.002535,
|
| 129 |
+
"S_rank": 0.09221231228343889,
|
| 130 |
+
"S_z": 0.11245578190554997
|
| 131 |
+
},
|
| 132 |
+
"VERDICT": "REJECT \u2014 THE OPERATORS ARE FUNCTIONALLY IDENTICAL",
|
| 133 |
+
"controls": {
|
| 134 |
+
"C1_zero_perturbation_is_noop": {
|
| 135 |
+
"required": "r=0 reproduces the unperturbed read exactly in both spaces",
|
| 136 |
+
"S_rank": 0.0,
|
| 137 |
+
"S_z": 0.0,
|
| 138 |
+
"fires": true
|
| 139 |
+
},
|
| 140 |
+
"C2_logit_anchor": {
|
| 141 |
+
"required": "logit_I flat-mean-7 min == 0.19810852520167826 within 1e-6",
|
| 142 |
+
"observed": 0.1981085192666858,
|
| 143 |
+
"abs_diff": 5.934992469702749e-09,
|
| 144 |
+
"fires": true
|
| 145 |
+
},
|
| 146 |
+
"C3_harness_responds": {
|
| 147 |
+
"required": "at r=5e-2 at least one space must move >= 1 seed SD, else the scoring path is not responding to the operator and every number is void",
|
| 148 |
+
"S_rank": 7.108817727719558,
|
| 149 |
+
"S_z": 5.607801800673574,
|
| 150 |
+
"fires": true
|
| 151 |
+
},
|
| 152 |
+
"C4_monotone_in_r": {
|
| 153 |
+
"required": "S_z non-decreasing across magnitudes",
|
| 154 |
+
"S_z_by_r": [
|
| 155 |
+
0.0,
|
| 156 |
+
4.9872383888545455e-06,
|
| 157 |
+
0.11245578190554997,
|
| 158 |
+
0.5224246102042763,
|
| 159 |
+
5.607801800673574
|
| 160 |
+
],
|
| 161 |
+
"fires": true
|
| 162 |
+
}
|
| 163 |
+
},
|
| 164 |
+
"declared_bias": "Gaussian noise is not a fitter difference. This bounds sensitivity to a perturbation of that size; it does not reproduce E66's specific operator delta, which is structural (accumulation order, batching, kernels).",
|
| 165 |
+
"e66_context": "results/e66_fitter_equivalence_cuda.json: trainval vs stored operator max_rel 2.535e-03 with read difference exactly 0.0; its C1 did not fire and its verdict flags the stored lens's provenance. The flagged file is the base operator used here and in CV3.",
|
| 166 |
+
"provenance": {
|
| 167 |
+
"schema_version": 1,
|
| 168 |
+
"experiment": "CV5",
|
| 169 |
+
"utc": "2026-08-30T00:18:45+00:00",
|
| 170 |
+
"script": {
|
| 171 |
+
"path": "experiments/cv5_metric_sensitivity.py",
|
| 172 |
+
"sha256": "f153b87305050c47829552d76859bc6d11c9b343419628b8122d2cc3d9c06ef6",
|
| 173 |
+
"n_lines": 198
|
| 174 |
+
},
|
| 175 |
+
"git": {
|
| 176 |
+
"commit": "c9c49697e295c0fec4225552836fc9733b279f75",
|
| 177 |
+
"commit_short": "c9c49697e295",
|
| 178 |
+
"git_repository_present": true,
|
| 179 |
+
"branch": "pythia",
|
| 180 |
+
"dirty": true,
|
| 181 |
+
"dirty_files": [
|
| 182 |
+
"CLAUDE.md",
|
| 183 |
+
"README.md",
|
| 184 |
+
"corpora/manifest_ood.json",
|
| 185 |
+
"docs/HANDOFF.md",
|
| 186 |
+
"docs/REPO_REVIEW_PROMPT.md",
|
| 187 |
+
"docs/context/AGGREGATION_POLICY.md",
|
| 188 |
+
"docs/context/CONTEXT.md",
|
| 189 |
+
"docs/context/IMPLEMENTATION_RULES.md",
|
| 190 |
+
"docs/context/STATE.md",
|
| 191 |
+
"docs/experiments/INDEX.md",
|
| 192 |
+
"docs/experiments/METRICS.md",
|
| 193 |
+
"docs/experiments/PROPOSED.md",
|
| 194 |
+
"docs/experiments/PROVENANCE.md",
|
| 195 |
+
"docs/experiments/descriptions/E64.md",
|
| 196 |
+
"docs/experiments/descriptions/EN1.md",
|
| 197 |
+
"docs/experiments/preregs/CV3_margins.md",
|
| 198 |
+
"docs/experiments/preregs/CV4_capability_ladder.md",
|
| 199 |
+
"docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 200 |
+
"docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 201 |
+
"docs/experiments/preregs/CV7_1b_rung.md",
|
| 202 |
+
"docs/experiments/preregs/DN1_design_nulls.md",
|
| 203 |
+
"docs/experiments/preregs/E33_logit_baseline.md",
|
| 204 |
+
"docs/experiments/preregs/E36_qladder.md",
|
| 205 |
+
"docs/experiments/preregs/E48_crossover.md",
|
| 206 |
+
"docs/experiments/preregs/E48b_containment.md",
|
| 207 |
+
"docs/experiments/preregs/E48c_exposure_vs_read.md",
|
| 208 |
+
"docs/experiments/preregs/E51_interaction.md",
|
| 209 |
+
"docs/experiments/preregs/E52_factorial.md",
|
| 210 |
+
"docs/experiments/preregs/E53_ladder_summary.md",
|
| 211 |
+
"docs/experiments/preregs/E55_matrix_robustness.md",
|
| 212 |
+
"docs/experiments/preregs/E56_predictor_registry.md",
|
| 213 |
+
"docs/experiments/preregs/E57_grid_variance_ci.md",
|
| 214 |
+
"docs/experiments/preregs/E58_algebra_audit.md",
|
| 215 |
+
"docs/experiments/preregs/E59_read_dose_response.md",
|
| 216 |
+
"docs/experiments/preregs/E60_fitter_determinism.md",
|
| 217 |
+
"docs/experiments/preregs/E61_randomized_network_null.md",
|
| 218 |
+
"docs/experiments/preregs/E62_ladder1b_corrected_band.md",
|
| 219 |
+
"docs/experiments/preregs/E65_training_axis_floor.md",
|
| 220 |
+
"docs/experiments/preregs/E66b_determinism_floor.md",
|
| 221 |
+
"docs/experiments/preregs/OS1_operator_space.md",
|
| 222 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md",
|
| 223 |
+
"docs/experiments/preregs/R1_grid_rstrip.md",
|
| 224 |
+
"docs/experiments/preregs/R2_e48_rstrip_arm.md",
|
| 225 |
+
"docs/experiments/preregs/R3_ladder410_cpu.md",
|
| 226 |
+
"docs/experiments/preregs/R4_corrections.md",
|
| 227 |
+
"docs/experiments/preregs/R4b_e36_rstrip.md",
|
| 228 |
+
"docs/experiments/preregs/R5_corpus_axis_uncertainty.md",
|
| 229 |
+
"docs/experiments/preregs/R6_within_source_resampling.md",
|
| 230 |
+
"docs/experiments/preregs/R8_ladder_rstrip_S2.md",
|
| 231 |
+
"docs/experiments/preregs/R9_permutation_calibrated_min.md",
|
| 232 |
+
"docs/experiments/preregs/superseded/PREREG_E48_CROSSOVER.md",
|
| 233 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 234 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 235 |
+
"docs/reproducibility/RELEASE.md",
|
| 236 |
+
"docs/reproducibility/REVIEWER_REPRO.md",
|
| 237 |
+
"docs/reproducibility/T1_SWEEP_SPEC.md",
|
| 238 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 239 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 240 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 241 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 242 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 243 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md",
|
| 244 |
+
"docs/reviews/REVIEW_INSIGHTS_2026-08-25.md",
|
| 245 |
+
"docs/validity/CONSTRUCT_VALIDITY.md"
|
| 246 |
+
],
|
| 247 |
+
"dirty_diff_sha256": "6b143c292229b3d2480d1b73243675f0db5e23b27f66e4215e0a9c1947a761a8",
|
| 248 |
+
"describe": "c9c4969-dirty"
|
| 249 |
+
},
|
| 250 |
+
"argv": [
|
| 251 |
+
"experiments/cv5_metric_sensitivity.py"
|
| 252 |
+
],
|
| 253 |
+
"cwd": ".",
|
| 254 |
+
"inputs": [
|
| 255 |
+
{
|
| 256 |
+
"path": "results/e48/lens_INSTREAM_Pile-CC_410m_n200_s0.pt",
|
| 257 |
+
"sha256": "4f00cc3d7450c3646b78f002e794572ca27d6ab175c70626d2d5fb7fccf28e20",
|
| 258 |
+
"bytes": 27268349,
|
| 259 |
+
"exists": true
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"path": "results/e66_fitter_equivalence_cuda.json",
|
| 263 |
+
"sha256": "d6e9bd3cd0c9e9426fc69e9d8d711f9a46e934214aaa52f67e877d947ec55cd2",
|
| 264 |
+
"bytes": 3208,
|
| 265 |
+
"exists": true
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"path": "results/cv3_margins_410m.json",
|
| 269 |
+
"sha256": "6cc200df5827b3c358d533934a37d892bfe1f7e9893bc4e4480421700afe7e99",
|
| 270 |
+
"bytes": 14659,
|
| 271 |
+
"exists": true
|
| 272 |
+
}
|
| 273 |
+
],
|
| 274 |
+
"env": {
|
| 275 |
+
"python": "3.14.2",
|
| 276 |
+
"torch": "2.13.0",
|
| 277 |
+
"transformers": "5.14.1",
|
| 278 |
+
"numpy": "2.5.1",
|
| 279 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 280 |
+
"machine": "arm64",
|
| 281 |
+
"hostname": "<redacted for double-blind review>",
|
| 282 |
+
"tf32": {
|
| 283 |
+
"allow_tf32_matmul": false,
|
| 284 |
+
"allow_tf32_cudnn": true,
|
| 285 |
+
"NVIDIA_TF32_OVERRIDE": null
|
| 286 |
+
},
|
| 287 |
+
"threads": {
|
| 288 |
+
"torch_num_threads": 8,
|
| 289 |
+
"torch_num_interop_threads": 14,
|
| 290 |
+
"OMP_NUM_THREADS": null,
|
| 291 |
+
"os_cpu_count": 14
|
| 292 |
+
}
|
| 293 |
+
},
|
| 294 |
+
"payload_sha256": "cde8769ac592f81096a80090b47a40f14c66626684029dad622f967b054ed1d9"
|
| 295 |
+
}
|
| 296 |
+
}
|
results/cv6/cv6_1.4b_n25.json
ADDED
|
@@ -0,0 +1,1336 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV6 \u2014 the per-family corpus contrast, one model",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 4 |
+
"status": "PRE-REGISTERED",
|
| 5 |
+
"decision_rule_verbatim": "Evaluated PER FAMILY; the verdict is the count across the five. REPLICATES if R(f,2.8B) >= 10 on >= 4 of 5 families AND ordering tau >= 0.6 on >= 3 of 5. ATTENUATES if R(f,2.8B) < 5 on >= 4 of 5 families. UNCLEAR otherwise, including a split where some families replicate and others do not \u2014 a split is an informative outcome, not a failure. Do not pool to force a verdict. Do not add models to chase a threshold. Do not drop a family.",
|
| 6 |
+
"model": "EleutherAI/pythia-1.4b-deduped",
|
| 7 |
+
"short": "1.4b",
|
| 8 |
+
"n_layers": 24,
|
| 9 |
+
"d_model": 2048,
|
| 10 |
+
"band": [
|
| 11 |
+
9,
|
| 12 |
+
10,
|
| 13 |
+
11,
|
| 14 |
+
12,
|
| 15 |
+
13,
|
| 16 |
+
14,
|
| 17 |
+
15,
|
| 18 |
+
16,
|
| 19 |
+
17,
|
| 20 |
+
18,
|
| 21 |
+
19,
|
| 22 |
+
20,
|
| 23 |
+
21
|
| 24 |
+
],
|
| 25 |
+
"band_rule": "floor(0.38*L)..floor(0.92*L), layers < L-2",
|
| 26 |
+
"band_matches_rule": true,
|
| 27 |
+
"N": 25,
|
| 28 |
+
"dim_batch": 128,
|
| 29 |
+
"max_seq_len": 128,
|
| 30 |
+
"skip_first": 16,
|
| 31 |
+
"target_layer_arg": -2,
|
| 32 |
+
"target_layer_effective": 22,
|
| 33 |
+
"device": "cuda",
|
| 34 |
+
"dtype": "float32",
|
| 35 |
+
"tf32": {
|
| 36 |
+
"matmul_allow_tf32": false,
|
| 37 |
+
"cudnn_allow_tf32": false,
|
| 38 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 39 |
+
},
|
| 40 |
+
"gpu": "NVIDIA H100 NVL",
|
| 41 |
+
"gpu_total_gb": 93.08551025390625,
|
| 42 |
+
"peak_vram_gb": 26.350069046020508,
|
| 43 |
+
"torch": "2.11.0+cu128",
|
| 44 |
+
"git_commit_at_launch": "518d192e9e215603957c85ac4a1308153bf4cf06",
|
| 45 |
+
"K": [
|
| 46 |
+
1,
|
| 47 |
+
2,
|
| 48 |
+
5,
|
| 49 |
+
10,
|
| 50 |
+
20,
|
| 51 |
+
50,
|
| 52 |
+
100
|
| 53 |
+
],
|
| 54 |
+
"admitted_sets": [
|
| 55 |
+
"multihop",
|
| 56 |
+
"multilingual",
|
| 57 |
+
"order-ops",
|
| 58 |
+
"poetry",
|
| 59 |
+
"typo"
|
| 60 |
+
],
|
| 61 |
+
"corpora": [
|
| 62 |
+
"Wikipedia_en",
|
| 63 |
+
"USPTO_Backgrounds",
|
| 64 |
+
"Pile-CC",
|
| 65 |
+
"StackExchange",
|
| 66 |
+
"Github"
|
| 67 |
+
],
|
| 68 |
+
"seeds": [
|
| 69 |
+
0,
|
| 70 |
+
1,
|
| 71 |
+
2
|
| 72 |
+
],
|
| 73 |
+
"readout_convention": "STRIPPED (corrected), no prefix, flat-mean-7",
|
| 74 |
+
"pooling_convention": "flat pool over (item, intermediate) pairs WITHIN a family; rank = min over band then flat-mean-7 pass@k; z = max over band then mean over pairs",
|
| 75 |
+
"n_items": 449,
|
| 76 |
+
"n_pairs": 801,
|
| 77 |
+
"n_pairs_per_family": {
|
| 78 |
+
"multihop": 103,
|
| 79 |
+
"multilingual": 394,
|
| 80 |
+
"order-ops": 110,
|
| 81 |
+
"poetry": 98,
|
| 82 |
+
"typo": 96
|
| 83 |
+
},
|
| 84 |
+
"per_arm": {
|
| 85 |
+
"logit_I": {
|
| 86 |
+
"multihop": {
|
| 87 |
+
"rank": 0.2496532593619972,
|
| 88 |
+
"z": 3.8530033965712613,
|
| 89 |
+
"n_pairs": 103
|
| 90 |
+
},
|
| 91 |
+
"multilingual": {
|
| 92 |
+
"rank": 0.1899927483683829,
|
| 93 |
+
"z": 3.2260746860564664,
|
| 94 |
+
"n_pairs": 394
|
| 95 |
+
},
|
| 96 |
+
"order-ops": {
|
| 97 |
+
"rank": 0.5597402597402598,
|
| 98 |
+
"z": 5.515925762870095,
|
| 99 |
+
"n_pairs": 110
|
| 100 |
+
},
|
| 101 |
+
"poetry": {
|
| 102 |
+
"rank": 0.008746355685131196,
|
| 103 |
+
"z": 2.4992136784962247,
|
| 104 |
+
"n_pairs": 98
|
| 105 |
+
},
|
| 106 |
+
"typo": {
|
| 107 |
+
"rank": 0.1889880952380952,
|
| 108 |
+
"z": 3.7020364006360373,
|
| 109 |
+
"n_pairs": 96
|
| 110 |
+
}
|
| 111 |
+
},
|
| 112 |
+
"J|Wikipedia_en|s0": {
|
| 113 |
+
"multihop": {
|
| 114 |
+
"rank": 0.3106796116504854,
|
| 115 |
+
"z": 3.8606277831549782,
|
| 116 |
+
"n_pairs": 103
|
| 117 |
+
},
|
| 118 |
+
"multilingual": {
|
| 119 |
+
"rank": 0.1892675852066715,
|
| 120 |
+
"z": 3.0605827324886614,
|
| 121 |
+
"n_pairs": 394
|
| 122 |
+
},
|
| 123 |
+
"order-ops": {
|
| 124 |
+
"rank": 0.5740259740259741,
|
| 125 |
+
"z": 5.3812003374099735,
|
| 126 |
+
"n_pairs": 110
|
| 127 |
+
},
|
| 128 |
+
"poetry": {
|
| 129 |
+
"rank": 0.026239067055393583,
|
| 130 |
+
"z": 2.3095518247205384,
|
| 131 |
+
"n_pairs": 98
|
| 132 |
+
},
|
| 133 |
+
"typo": {
|
| 134 |
+
"rank": 0.3392857142857143,
|
| 135 |
+
"z": 3.6630871382852397,
|
| 136 |
+
"n_pairs": 96
|
| 137 |
+
}
|
| 138 |
+
},
|
| 139 |
+
"J|Wikipedia_en|s1": {
|
| 140 |
+
"multihop": {
|
| 141 |
+
"rank": 0.3092926490984743,
|
| 142 |
+
"z": 3.8443020744231142,
|
| 143 |
+
"n_pairs": 103
|
| 144 |
+
},
|
| 145 |
+
"multilingual": {
|
| 146 |
+
"rank": 0.1907179115300943,
|
| 147 |
+
"z": 3.077618185033653,
|
| 148 |
+
"n_pairs": 394
|
| 149 |
+
},
|
| 150 |
+
"order-ops": {
|
| 151 |
+
"rank": 0.5545454545454546,
|
| 152 |
+
"z": 5.271360293301669,
|
| 153 |
+
"n_pairs": 110
|
| 154 |
+
},
|
| 155 |
+
"poetry": {
|
| 156 |
+
"rank": 0.018950437317784258,
|
| 157 |
+
"z": 2.327480596547224,
|
| 158 |
+
"n_pairs": 98
|
| 159 |
+
},
|
| 160 |
+
"typo": {
|
| 161 |
+
"rank": 0.33779761904761907,
|
| 162 |
+
"z": 3.6704573507110276,
|
| 163 |
+
"n_pairs": 96
|
| 164 |
+
}
|
| 165 |
+
},
|
| 166 |
+
"J|Wikipedia_en|s2": {
|
| 167 |
+
"multihop": {
|
| 168 |
+
"rank": 0.3065187239944521,
|
| 169 |
+
"z": 3.843342467419152,
|
| 170 |
+
"n_pairs": 103
|
| 171 |
+
},
|
| 172 |
+
"multilingual": {
|
| 173 |
+
"rank": 0.1787527193618564,
|
| 174 |
+
"z": 3.0286048066797595,
|
| 175 |
+
"n_pairs": 394
|
| 176 |
+
},
|
| 177 |
+
"order-ops": {
|
| 178 |
+
"rank": 0.5610389610389611,
|
| 179 |
+
"z": 5.311658326062289,
|
| 180 |
+
"n_pairs": 110
|
| 181 |
+
},
|
| 182 |
+
"poetry": {
|
| 183 |
+
"rank": 0.021865889212827987,
|
| 184 |
+
"z": 2.32845254516115,
|
| 185 |
+
"n_pairs": 98
|
| 186 |
+
},
|
| 187 |
+
"typo": {
|
| 188 |
+
"rank": 0.33482142857142855,
|
| 189 |
+
"z": 3.67272778848807,
|
| 190 |
+
"n_pairs": 96
|
| 191 |
+
}
|
| 192 |
+
},
|
| 193 |
+
"J|USPTO_Backgrounds|s0": {
|
| 194 |
+
"multihop": {
|
| 195 |
+
"rank": 0.32177531206657417,
|
| 196 |
+
"z": 3.874361871515663,
|
| 197 |
+
"n_pairs": 103
|
| 198 |
+
},
|
| 199 |
+
"multilingual": {
|
| 200 |
+
"rank": 0.1783901377810007,
|
| 201 |
+
"z": 3.081772235746916,
|
| 202 |
+
"n_pairs": 394
|
| 203 |
+
},
|
| 204 |
+
"order-ops": {
|
| 205 |
+
"rank": 0.5727272727272729,
|
| 206 |
+
"z": 5.26822690855373,
|
| 207 |
+
"n_pairs": 110
|
| 208 |
+
},
|
| 209 |
+
"poetry": {
|
| 210 |
+
"rank": 0.021865889212827987,
|
| 211 |
+
"z": 2.352261223355118,
|
| 212 |
+
"n_pairs": 98
|
| 213 |
+
},
|
| 214 |
+
"typo": {
|
| 215 |
+
"rank": 0.4404761904761905,
|
| 216 |
+
"z": 3.9461715444922447,
|
| 217 |
+
"n_pairs": 96
|
| 218 |
+
}
|
| 219 |
+
},
|
| 220 |
+
"J|USPTO_Backgrounds|s1": {
|
| 221 |
+
"multihop": {
|
| 222 |
+
"rank": 0.319001386962552,
|
| 223 |
+
"z": 3.9217262459032742,
|
| 224 |
+
"n_pairs": 103
|
| 225 |
+
},
|
| 226 |
+
"multilingual": {
|
| 227 |
+
"rank": 0.1787527193618564,
|
| 228 |
+
"z": 3.1000644785498603,
|
| 229 |
+
"n_pairs": 394
|
| 230 |
+
},
|
| 231 |
+
"order-ops": {
|
| 232 |
+
"rank": 0.596103896103896,
|
| 233 |
+
"z": 5.3440335826440295,
|
| 234 |
+
"n_pairs": 110
|
| 235 |
+
},
|
| 236 |
+
"poetry": {
|
| 237 |
+
"rank": 0.021865889212827987,
|
| 238 |
+
"z": 2.3861058938259982,
|
| 239 |
+
"n_pairs": 98
|
| 240 |
+
},
|
| 241 |
+
"typo": {
|
| 242 |
+
"rank": 0.45238095238095244,
|
| 243 |
+
"z": 3.9819413671890893,
|
| 244 |
+
"n_pairs": 96
|
| 245 |
+
}
|
| 246 |
+
},
|
| 247 |
+
"J|USPTO_Backgrounds|s2": {
|
| 248 |
+
"multihop": {
|
| 249 |
+
"rank": 0.32593619972260746,
|
| 250 |
+
"z": 3.934496842541741,
|
| 251 |
+
"n_pairs": 103
|
| 252 |
+
},
|
| 253 |
+
"multilingual": {
|
| 254 |
+
"rank": 0.18419144307469182,
|
| 255 |
+
"z": 3.1254321258080187,
|
| 256 |
+
"n_pairs": 394
|
| 257 |
+
},
|
| 258 |
+
"order-ops": {
|
| 259 |
+
"rank": 0.5805194805194805,
|
| 260 |
+
"z": 5.343620314381339,
|
| 261 |
+
"n_pairs": 110
|
| 262 |
+
},
|
| 263 |
+
"poetry": {
|
| 264 |
+
"rank": 0.026239067055393583,
|
| 265 |
+
"z": 2.369717459897606,
|
| 266 |
+
"n_pairs": 98
|
| 267 |
+
},
|
| 268 |
+
"typo": {
|
| 269 |
+
"rank": 0.4568452380952381,
|
| 270 |
+
"z": 4.044803236921628,
|
| 271 |
+
"n_pairs": 96
|
| 272 |
+
}
|
| 273 |
+
},
|
| 274 |
+
"J|Pile-CC|s0": {
|
| 275 |
+
"multihop": {
|
| 276 |
+
"rank": 0.3203883495145631,
|
| 277 |
+
"z": 3.9831742198721876,
|
| 278 |
+
"n_pairs": 103
|
| 279 |
+
},
|
| 280 |
+
"multilingual": {
|
| 281 |
+
"rank": 0.20123277737490936,
|
| 282 |
+
"z": 3.2589381551984604,
|
| 283 |
+
"n_pairs": 394
|
| 284 |
+
},
|
| 285 |
+
"order-ops": {
|
| 286 |
+
"rank": 0.5870129870129871,
|
| 287 |
+
"z": 5.469181368567726,
|
| 288 |
+
"n_pairs": 110
|
| 289 |
+
},
|
| 290 |
+
"poetry": {
|
| 291 |
+
"rank": 0.018950437317784254,
|
| 292 |
+
"z": 2.397173393137601,
|
| 293 |
+
"n_pairs": 98
|
| 294 |
+
},
|
| 295 |
+
"typo": {
|
| 296 |
+
"rank": 0.3988095238095238,
|
| 297 |
+
"z": 4.014354603985946,
|
| 298 |
+
"n_pairs": 96
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"J|Pile-CC|s1": {
|
| 302 |
+
"multihop": {
|
| 303 |
+
"rank": 0.32593619972260746,
|
| 304 |
+
"z": 4.029289348032868,
|
| 305 |
+
"n_pairs": 103
|
| 306 |
+
},
|
| 307 |
+
"multilingual": {
|
| 308 |
+
"rank": 0.19361856417693982,
|
| 309 |
+
"z": 3.188464311627567,
|
| 310 |
+
"n_pairs": 394
|
| 311 |
+
},
|
| 312 |
+
"order-ops": {
|
| 313 |
+
"rank": 0.5922077922077923,
|
| 314 |
+
"z": 5.454147653146223,
|
| 315 |
+
"n_pairs": 110
|
| 316 |
+
},
|
| 317 |
+
"poetry": {
|
| 318 |
+
"rank": 0.01749271137026239,
|
| 319 |
+
"z": 2.3707593989615536,
|
| 320 |
+
"n_pairs": 98
|
| 321 |
+
},
|
| 322 |
+
"typo": {
|
| 323 |
+
"rank": 0.4107142857142857,
|
| 324 |
+
"z": 4.036523262659709,
|
| 325 |
+
"n_pairs": 96
|
| 326 |
+
}
|
| 327 |
+
},
|
| 328 |
+
"J|Pile-CC|s2": {
|
| 329 |
+
"multihop": {
|
| 330 |
+
"rank": 0.32177531206657417,
|
| 331 |
+
"z": 3.953363766947996,
|
| 332 |
+
"n_pairs": 103
|
| 333 |
+
},
|
| 334 |
+
"multilingual": {
|
| 335 |
+
"rank": 0.1943437273386512,
|
| 336 |
+
"z": 3.190489783202331,
|
| 337 |
+
"n_pairs": 394
|
| 338 |
+
},
|
| 339 |
+
"order-ops": {
|
| 340 |
+
"rank": 0.5870129870129871,
|
| 341 |
+
"z": 5.500762580741536,
|
| 342 |
+
"n_pairs": 110
|
| 343 |
+
},
|
| 344 |
+
"poetry": {
|
| 345 |
+
"rank": 0.013119533527696792,
|
| 346 |
+
"z": 2.357462505296785,
|
| 347 |
+
"n_pairs": 98
|
| 348 |
+
},
|
| 349 |
+
"typo": {
|
| 350 |
+
"rank": 0.4047619047619047,
|
| 351 |
+
"z": 4.010696103175481,
|
| 352 |
+
"n_pairs": 96
|
| 353 |
+
}
|
| 354 |
+
},
|
| 355 |
+
"J|StackExchange|s0": {
|
| 356 |
+
"multihop": {
|
| 357 |
+
"rank": 0.3883495145631068,
|
| 358 |
+
"z": 4.0767282235969615,
|
| 359 |
+
"n_pairs": 103
|
| 360 |
+
},
|
| 361 |
+
"multilingual": {
|
| 362 |
+
"rank": 0.21537345902828137,
|
| 363 |
+
"z": 3.1823133939716417,
|
| 364 |
+
"n_pairs": 394
|
| 365 |
+
},
|
| 366 |
+
"order-ops": {
|
| 367 |
+
"rank": 0.5935064935064934,
|
| 368 |
+
"z": 5.240476467392662,
|
| 369 |
+
"n_pairs": 110
|
| 370 |
+
},
|
| 371 |
+
"poetry": {
|
| 372 |
+
"rank": 0.021865889212827987,
|
| 373 |
+
"z": 2.3227810050759996,
|
| 374 |
+
"n_pairs": 98
|
| 375 |
+
},
|
| 376 |
+
"typo": {
|
| 377 |
+
"rank": 0.46428571428571425,
|
| 378 |
+
"z": 4.006702713668346,
|
| 379 |
+
"n_pairs": 96
|
| 380 |
+
}
|
| 381 |
+
},
|
| 382 |
+
"J|StackExchange|s1": {
|
| 383 |
+
"multihop": {
|
| 384 |
+
"rank": 0.391123439667129,
|
| 385 |
+
"z": 4.067388656069931,
|
| 386 |
+
"n_pairs": 103
|
| 387 |
+
},
|
| 388 |
+
"multilingual": {
|
| 389 |
+
"rank": 0.22588832487309643,
|
| 390 |
+
"z": 3.2193740812049905,
|
| 391 |
+
"n_pairs": 394
|
| 392 |
+
},
|
| 393 |
+
"order-ops": {
|
| 394 |
+
"rank": 0.5844155844155844,
|
| 395 |
+
"z": 5.164196674390273,
|
| 396 |
+
"n_pairs": 110
|
| 397 |
+
},
|
| 398 |
+
"poetry": {
|
| 399 |
+
"rank": 0.018950437317784254,
|
| 400 |
+
"z": 2.2825519923044713,
|
| 401 |
+
"n_pairs": 98
|
| 402 |
+
},
|
| 403 |
+
"typo": {
|
| 404 |
+
"rank": 0.45982142857142855,
|
| 405 |
+
"z": 3.988087110221386,
|
| 406 |
+
"n_pairs": 96
|
| 407 |
+
}
|
| 408 |
+
},
|
| 409 |
+
"J|StackExchange|s2": {
|
| 410 |
+
"multihop": {
|
| 411 |
+
"rank": 0.4105409153952843,
|
| 412 |
+
"z": 4.105738782188268,
|
| 413 |
+
"n_pairs": 103
|
| 414 |
+
},
|
| 415 |
+
"multilingual": {
|
| 416 |
+
"rank": 0.2411167512690355,
|
| 417 |
+
"z": 3.290932434767031,
|
| 418 |
+
"n_pairs": 394
|
| 419 |
+
},
|
| 420 |
+
"order-ops": {
|
| 421 |
+
"rank": 0.5909090909090909,
|
| 422 |
+
"z": 5.2581051761453805,
|
| 423 |
+
"n_pairs": 110
|
| 424 |
+
},
|
| 425 |
+
"poetry": {
|
| 426 |
+
"rank": 0.021865889212827987,
|
| 427 |
+
"z": 2.309936959524544,
|
| 428 |
+
"n_pairs": 98
|
| 429 |
+
},
|
| 430 |
+
"typo": {
|
| 431 |
+
"rank": 0.4613095238095238,
|
| 432 |
+
"z": 3.990582413971424,
|
| 433 |
+
"n_pairs": 96
|
| 434 |
+
}
|
| 435 |
+
},
|
| 436 |
+
"J|Github|s0": {
|
| 437 |
+
"multihop": {
|
| 438 |
+
"rank": 0.30374479889042993,
|
| 439 |
+
"z": 3.6554290958978597,
|
| 440 |
+
"n_pairs": 103
|
| 441 |
+
},
|
| 442 |
+
"multilingual": {
|
| 443 |
+
"rank": 0.13379260333575055,
|
| 444 |
+
"z": 2.789235497035351,
|
| 445 |
+
"n_pairs": 394
|
| 446 |
+
},
|
| 447 |
+
"order-ops": {
|
| 448 |
+
"rank": 0.5571428571428572,
|
| 449 |
+
"z": 5.291046177257191,
|
| 450 |
+
"n_pairs": 110
|
| 451 |
+
},
|
| 452 |
+
"poetry": {
|
| 453 |
+
"rank": 0.011661807580174929,
|
| 454 |
+
"z": 2.0492812370767397,
|
| 455 |
+
"n_pairs": 98
|
| 456 |
+
},
|
| 457 |
+
"typo": {
|
| 458 |
+
"rank": 0.20833333333333334,
|
| 459 |
+
"z": 3.194101763268312,
|
| 460 |
+
"n_pairs": 96
|
| 461 |
+
}
|
| 462 |
+
},
|
| 463 |
+
"J|Github|s1": {
|
| 464 |
+
"multihop": {
|
| 465 |
+
"rank": 0.32871012482662965,
|
| 466 |
+
"z": 3.81565054412027,
|
| 467 |
+
"n_pairs": 103
|
| 468 |
+
},
|
| 469 |
+
"multilingual": {
|
| 470 |
+
"rank": 0.16678752719361858,
|
| 471 |
+
"z": 2.9543404418804924,
|
| 472 |
+
"n_pairs": 394
|
| 473 |
+
},
|
| 474 |
+
"order-ops": {
|
| 475 |
+
"rank": 0.5844155844155844,
|
| 476 |
+
"z": 5.422733649340543,
|
| 477 |
+
"n_pairs": 110
|
| 478 |
+
},
|
| 479 |
+
"poetry": {
|
| 480 |
+
"rank": 0.013119533527696792,
|
| 481 |
+
"z": 2.1037646602002944,
|
| 482 |
+
"n_pairs": 98
|
| 483 |
+
},
|
| 484 |
+
"typo": {
|
| 485 |
+
"rank": 0.2574404761904762,
|
| 486 |
+
"z": 3.366413981964191,
|
| 487 |
+
"n_pairs": 96
|
| 488 |
+
}
|
| 489 |
+
},
|
| 490 |
+
"J|Github|s2": {
|
| 491 |
+
"multihop": {
|
| 492 |
+
"rank": 0.3398058252427184,
|
| 493 |
+
"z": 3.787343767082807,
|
| 494 |
+
"n_pairs": 103
|
| 495 |
+
},
|
| 496 |
+
"multilingual": {
|
| 497 |
+
"rank": 0.14176939811457578,
|
| 498 |
+
"z": 2.774672096755904,
|
| 499 |
+
"n_pairs": 394
|
| 500 |
+
},
|
| 501 |
+
"order-ops": {
|
| 502 |
+
"rank": 0.5701298701298702,
|
| 503 |
+
"z": 5.435313402522694,
|
| 504 |
+
"n_pairs": 110
|
| 505 |
+
},
|
| 506 |
+
"poetry": {
|
| 507 |
+
"rank": 0.01603498542274052,
|
| 508 |
+
"z": 2.0549783214014403,
|
| 509 |
+
"n_pairs": 98
|
| 510 |
+
},
|
| 511 |
+
"typo": {
|
| 512 |
+
"rank": 0.22470238095238093,
|
| 513 |
+
"z": 3.2467485206822553,
|
| 514 |
+
"n_pairs": 96
|
| 515 |
+
}
|
| 516 |
+
}
|
| 517 |
+
},
|
| 518 |
+
"per_arm_fp16_roundtrip": {
|
| 519 |
+
"J|Wikipedia_en|s0": {
|
| 520 |
+
"multihop": {
|
| 521 |
+
"rank": 0.3106796116504854,
|
| 522 |
+
"z": 3.8606872315545684,
|
| 523 |
+
"n_pairs": 103
|
| 524 |
+
},
|
| 525 |
+
"multilingual": {
|
| 526 |
+
"rank": 0.1892675852066715,
|
| 527 |
+
"z": 3.060567912413989,
|
| 528 |
+
"n_pairs": 394
|
| 529 |
+
},
|
| 530 |
+
"order-ops": {
|
| 531 |
+
"rank": 0.5740259740259741,
|
| 532 |
+
"z": 5.381180027398196,
|
| 533 |
+
"n_pairs": 110
|
| 534 |
+
},
|
| 535 |
+
"poetry": {
|
| 536 |
+
"rank": 0.026239067055393583,
|
| 537 |
+
"z": 2.309572807988342,
|
| 538 |
+
"n_pairs": 98
|
| 539 |
+
},
|
| 540 |
+
"typo": {
|
| 541 |
+
"rank": 0.3392857142857143,
|
| 542 |
+
"z": 3.6631390204032264,
|
| 543 |
+
"n_pairs": 96
|
| 544 |
+
}
|
| 545 |
+
},
|
| 546 |
+
"J|Wikipedia_en|s1": {
|
| 547 |
+
"multihop": {
|
| 548 |
+
"rank": 0.3092926490984743,
|
| 549 |
+
"z": 3.844225617288386,
|
| 550 |
+
"n_pairs": 103
|
| 551 |
+
},
|
| 552 |
+
"multilingual": {
|
| 553 |
+
"rank": 0.1907179115300943,
|
| 554 |
+
"z": 3.0775656076857283,
|
| 555 |
+
"n_pairs": 394
|
| 556 |
+
},
|
| 557 |
+
"order-ops": {
|
| 558 |
+
"rank": 0.5545454545454546,
|
| 559 |
+
"z": 5.271341731331566,
|
| 560 |
+
"n_pairs": 110
|
| 561 |
+
},
|
| 562 |
+
"poetry": {
|
| 563 |
+
"rank": 0.018950437317784258,
|
| 564 |
+
"z": 2.3274755514397913,
|
| 565 |
+
"n_pairs": 98
|
| 566 |
+
},
|
| 567 |
+
"typo": {
|
| 568 |
+
"rank": 0.33779761904761907,
|
| 569 |
+
"z": 3.6704695199926696,
|
| 570 |
+
"n_pairs": 96
|
| 571 |
+
}
|
| 572 |
+
},
|
| 573 |
+
"J|Wikipedia_en|s2": {
|
| 574 |
+
"multihop": {
|
| 575 |
+
"rank": 0.30513176144244103,
|
| 576 |
+
"z": 3.8434583180159043,
|
| 577 |
+
"n_pairs": 103
|
| 578 |
+
},
|
| 579 |
+
"multilingual": {
|
| 580 |
+
"rank": 0.1787527193618564,
|
| 581 |
+
"z": 3.0286072580342367,
|
| 582 |
+
"n_pairs": 394
|
| 583 |
+
},
|
| 584 |
+
"order-ops": {
|
| 585 |
+
"rank": 0.5610389610389611,
|
| 586 |
+
"z": 5.311677533929998,
|
| 587 |
+
"n_pairs": 110
|
| 588 |
+
},
|
| 589 |
+
"poetry": {
|
| 590 |
+
"rank": 0.021865889212827987,
|
| 591 |
+
"z": 2.328465061528342,
|
| 592 |
+
"n_pairs": 98
|
| 593 |
+
},
|
| 594 |
+
"typo": {
|
| 595 |
+
"rank": 0.33482142857142855,
|
| 596 |
+
"z": 3.6727247970799604,
|
| 597 |
+
"n_pairs": 96
|
| 598 |
+
}
|
| 599 |
+
},
|
| 600 |
+
"J|USPTO_Backgrounds|s0": {
|
| 601 |
+
"multihop": {
|
| 602 |
+
"rank": 0.32177531206657417,
|
| 603 |
+
"z": 3.8744102160907485,
|
| 604 |
+
"n_pairs": 103
|
| 605 |
+
},
|
| 606 |
+
"multilingual": {
|
| 607 |
+
"rank": 0.1783901377810007,
|
| 608 |
+
"z": 3.081757281486153,
|
| 609 |
+
"n_pairs": 394
|
| 610 |
+
},
|
| 611 |
+
"order-ops": {
|
| 612 |
+
"rank": 0.5727272727272729,
|
| 613 |
+
"z": 5.268081293322823,
|
| 614 |
+
"n_pairs": 110
|
| 615 |
+
},
|
| 616 |
+
"poetry": {
|
| 617 |
+
"rank": 0.021865889212827987,
|
| 618 |
+
"z": 2.35228185568537,
|
| 619 |
+
"n_pairs": 98
|
| 620 |
+
},
|
| 621 |
+
"typo": {
|
| 622 |
+
"rank": 0.4404761904761905,
|
| 623 |
+
"z": 3.946172351638476,
|
| 624 |
+
"n_pairs": 96
|
| 625 |
+
}
|
| 626 |
+
},
|
| 627 |
+
"J|USPTO_Backgrounds|s1": {
|
| 628 |
+
"multihop": {
|
| 629 |
+
"rank": 0.3203883495145631,
|
| 630 |
+
"z": 3.921612312492815,
|
| 631 |
+
"n_pairs": 103
|
| 632 |
+
},
|
| 633 |
+
"multilingual": {
|
| 634 |
+
"rank": 0.1783901377810007,
|
| 635 |
+
"z": 3.100029178379756,
|
| 636 |
+
"n_pairs": 394
|
| 637 |
+
},
|
| 638 |
+
"order-ops": {
|
| 639 |
+
"rank": 0.596103896103896,
|
| 640 |
+
"z": 5.34396392323754,
|
| 641 |
+
"n_pairs": 110
|
| 642 |
+
},
|
| 643 |
+
"poetry": {
|
| 644 |
+
"rank": 0.021865889212827987,
|
| 645 |
+
"z": 2.386144852151676,
|
| 646 |
+
"n_pairs": 98
|
| 647 |
+
},
|
| 648 |
+
"typo": {
|
| 649 |
+
"rank": 0.45238095238095244,
|
| 650 |
+
"z": 3.9819253981113434,
|
| 651 |
+
"n_pairs": 96
|
| 652 |
+
}
|
| 653 |
+
},
|
| 654 |
+
"J|USPTO_Backgrounds|s2": {
|
| 655 |
+
"multihop": {
|
| 656 |
+
"rank": 0.32871012482662965,
|
| 657 |
+
"z": 3.934536116794475,
|
| 658 |
+
"n_pairs": 103
|
| 659 |
+
},
|
| 660 |
+
"multilingual": {
|
| 661 |
+
"rank": 0.18419144307469182,
|
| 662 |
+
"z": 3.125436220374809,
|
| 663 |
+
"n_pairs": 394
|
| 664 |
+
},
|
| 665 |
+
"order-ops": {
|
| 666 |
+
"rank": 0.5805194805194805,
|
| 667 |
+
"z": 5.34366934299469,
|
| 668 |
+
"n_pairs": 110
|
| 669 |
+
},
|
| 670 |
+
"poetry": {
|
| 671 |
+
"rank": 0.026239067055393583,
|
| 672 |
+
"z": 2.3696852350721556,
|
| 673 |
+
"n_pairs": 98
|
| 674 |
+
},
|
| 675 |
+
"typo": {
|
| 676 |
+
"rank": 0.4568452380952381,
|
| 677 |
+
"z": 4.044789165258408,
|
| 678 |
+
"n_pairs": 96
|
| 679 |
+
}
|
| 680 |
+
},
|
| 681 |
+
"J|Pile-CC|s0": {
|
| 682 |
+
"multihop": {
|
| 683 |
+
"rank": 0.3203883495145631,
|
| 684 |
+
"z": 3.9830801978851986,
|
| 685 |
+
"n_pairs": 103
|
| 686 |
+
},
|
| 687 |
+
"multilingual": {
|
| 688 |
+
"rank": 0.20123277737490936,
|
| 689 |
+
"z": 3.258956805249761,
|
| 690 |
+
"n_pairs": 394
|
| 691 |
+
},
|
| 692 |
+
"order-ops": {
|
| 693 |
+
"rank": 0.5870129870129871,
|
| 694 |
+
"z": 5.469027082486586,
|
| 695 |
+
"n_pairs": 110
|
| 696 |
+
},
|
| 697 |
+
"poetry": {
|
| 698 |
+
"rank": 0.018950437317784254,
|
| 699 |
+
"z": 2.3971886932849884,
|
| 700 |
+
"n_pairs": 98
|
| 701 |
+
},
|
| 702 |
+
"typo": {
|
| 703 |
+
"rank": 0.3988095238095238,
|
| 704 |
+
"z": 4.014335438609123,
|
| 705 |
+
"n_pairs": 96
|
| 706 |
+
}
|
| 707 |
+
},
|
| 708 |
+
"J|Pile-CC|s1": {
|
| 709 |
+
"multihop": {
|
| 710 |
+
"rank": 0.32593619972260746,
|
| 711 |
+
"z": 4.029135812833471,
|
| 712 |
+
"n_pairs": 103
|
| 713 |
+
},
|
| 714 |
+
"multilingual": {
|
| 715 |
+
"rank": 0.19361856417693982,
|
| 716 |
+
"z": 3.188446533105095,
|
| 717 |
+
"n_pairs": 394
|
| 718 |
+
},
|
| 719 |
+
"order-ops": {
|
| 720 |
+
"rank": 0.5922077922077923,
|
| 721 |
+
"z": 5.454229792681607,
|
| 722 |
+
"n_pairs": 110
|
| 723 |
+
},
|
| 724 |
+
"poetry": {
|
| 725 |
+
"rank": 0.01749271137026239,
|
| 726 |
+
"z": 2.370773496068254,
|
| 727 |
+
"n_pairs": 98
|
| 728 |
+
},
|
| 729 |
+
"typo": {
|
| 730 |
+
"rank": 0.4092261904761905,
|
| 731 |
+
"z": 4.0365254531304045,
|
| 732 |
+
"n_pairs": 96
|
| 733 |
+
}
|
| 734 |
+
},
|
| 735 |
+
"J|Pile-CC|s2": {
|
| 736 |
+
"multihop": {
|
| 737 |
+
"rank": 0.32177531206657417,
|
| 738 |
+
"z": 3.9533760107836677,
|
| 739 |
+
"n_pairs": 103
|
| 740 |
+
},
|
| 741 |
+
"multilingual": {
|
| 742 |
+
"rank": 0.19398114575779551,
|
| 743 |
+
"z": 3.1904335273098825,
|
| 744 |
+
"n_pairs": 394
|
| 745 |
+
},
|
| 746 |
+
"order-ops": {
|
| 747 |
+
"rank": 0.5870129870129871,
|
| 748 |
+
"z": 5.500757904486223,
|
| 749 |
+
"n_pairs": 110
|
| 750 |
+
},
|
| 751 |
+
"poetry": {
|
| 752 |
+
"rank": 0.013119533527696792,
|
| 753 |
+
"z": 2.357441063438143,
|
| 754 |
+
"n_pairs": 98
|
| 755 |
+
},
|
| 756 |
+
"typo": {
|
| 757 |
+
"rank": 0.4047619047619047,
|
| 758 |
+
"z": 4.010686029990514,
|
| 759 |
+
"n_pairs": 96
|
| 760 |
+
}
|
| 761 |
+
},
|
| 762 |
+
"J|StackExchange|s0": {
|
| 763 |
+
"multihop": {
|
| 764 |
+
"rank": 0.3883495145631068,
|
| 765 |
+
"z": 4.076668833065959,
|
| 766 |
+
"n_pairs": 103
|
| 767 |
+
},
|
| 768 |
+
"multilingual": {
|
| 769 |
+
"rank": 0.21537345902828137,
|
| 770 |
+
"z": 3.1823146536870657,
|
| 771 |
+
"n_pairs": 394
|
| 772 |
+
},
|
| 773 |
+
"order-ops": {
|
| 774 |
+
"rank": 0.5935064935064934,
|
| 775 |
+
"z": 5.24048402634534,
|
| 776 |
+
"n_pairs": 110
|
| 777 |
+
},
|
| 778 |
+
"poetry": {
|
| 779 |
+
"rank": 0.021865889212827987,
|
| 780 |
+
"z": 2.3227494511069082,
|
| 781 |
+
"n_pairs": 98
|
| 782 |
+
},
|
| 783 |
+
"typo": {
|
| 784 |
+
"rank": 0.46428571428571425,
|
| 785 |
+
"z": 4.006694952646892,
|
| 786 |
+
"n_pairs": 96
|
| 787 |
+
}
|
| 788 |
+
},
|
| 789 |
+
"J|StackExchange|s1": {
|
| 790 |
+
"multihop": {
|
| 791 |
+
"rank": 0.391123439667129,
|
| 792 |
+
"z": 4.067319025113744,
|
| 793 |
+
"n_pairs": 103
|
| 794 |
+
},
|
| 795 |
+
"multilingual": {
|
| 796 |
+
"rank": 0.22588832487309643,
|
| 797 |
+
"z": 3.21935731233074,
|
| 798 |
+
"n_pairs": 394
|
| 799 |
+
},
|
| 800 |
+
"order-ops": {
|
| 801 |
+
"rank": 0.5844155844155844,
|
| 802 |
+
"z": 5.164137942140752,
|
| 803 |
+
"n_pairs": 110
|
| 804 |
+
},
|
| 805 |
+
"poetry": {
|
| 806 |
+
"rank": 0.018950437317784254,
|
| 807 |
+
"z": 2.2825479994014817,
|
| 808 |
+
"n_pairs": 98
|
| 809 |
+
},
|
| 810 |
+
"typo": {
|
| 811 |
+
"rank": 0.45982142857142855,
|
| 812 |
+
"z": 3.9880845795075097,
|
| 813 |
+
"n_pairs": 96
|
| 814 |
+
}
|
| 815 |
+
},
|
| 816 |
+
"J|StackExchange|s2": {
|
| 817 |
+
"multihop": {
|
| 818 |
+
"rank": 0.4091539528432732,
|
| 819 |
+
"z": 4.10576043140541,
|
| 820 |
+
"n_pairs": 103
|
| 821 |
+
},
|
| 822 |
+
"multilingual": {
|
| 823 |
+
"rank": 0.2411167512690355,
|
| 824 |
+
"z": 3.2908637373883107,
|
| 825 |
+
"n_pairs": 394
|
| 826 |
+
},
|
| 827 |
+
"order-ops": {
|
| 828 |
+
"rank": 0.5909090909090909,
|
| 829 |
+
"z": 5.258090955560857,
|
| 830 |
+
"n_pairs": 110
|
| 831 |
+
},
|
| 832 |
+
"poetry": {
|
| 833 |
+
"rank": 0.021865889212827987,
|
| 834 |
+
"z": 2.3098818407374986,
|
| 835 |
+
"n_pairs": 98
|
| 836 |
+
},
|
| 837 |
+
"typo": {
|
| 838 |
+
"rank": 0.4613095238095238,
|
| 839 |
+
"z": 3.9905417760213218,
|
| 840 |
+
"n_pairs": 96
|
| 841 |
+
}
|
| 842 |
+
},
|
| 843 |
+
"J|Github|s0": {
|
| 844 |
+
"multihop": {
|
| 845 |
+
"rank": 0.30374479889042993,
|
| 846 |
+
"z": 3.6555041063179092,
|
| 847 |
+
"n_pairs": 103
|
| 848 |
+
},
|
| 849 |
+
"multilingual": {
|
| 850 |
+
"rank": 0.13379260333575055,
|
| 851 |
+
"z": 2.7891979518578136,
|
| 852 |
+
"n_pairs": 394
|
| 853 |
+
},
|
| 854 |
+
"order-ops": {
|
| 855 |
+
"rank": 0.5571428571428572,
|
| 856 |
+
"z": 5.291175675392151,
|
| 857 |
+
"n_pairs": 110
|
| 858 |
+
},
|
| 859 |
+
"poetry": {
|
| 860 |
+
"rank": 0.011661807580174929,
|
| 861 |
+
"z": 2.0492619884257413,
|
| 862 |
+
"n_pairs": 98
|
| 863 |
+
},
|
| 864 |
+
"typo": {
|
| 865 |
+
"rank": 0.20833333333333334,
|
| 866 |
+
"z": 3.1940517015755177,
|
| 867 |
+
"n_pairs": 96
|
| 868 |
+
}
|
| 869 |
+
},
|
| 870 |
+
"J|Github|s1": {
|
| 871 |
+
"multihop": {
|
| 872 |
+
"rank": 0.32871012482662965,
|
| 873 |
+
"z": 3.815705211417189,
|
| 874 |
+
"n_pairs": 103
|
| 875 |
+
},
|
| 876 |
+
"multilingual": {
|
| 877 |
+
"rank": 0.16678752719361858,
|
| 878 |
+
"z": 2.954350498122007,
|
| 879 |
+
"n_pairs": 394
|
| 880 |
+
},
|
| 881 |
+
"order-ops": {
|
| 882 |
+
"rank": 0.5844155844155844,
|
| 883 |
+
"z": 5.422900062257593,
|
| 884 |
+
"n_pairs": 110
|
| 885 |
+
},
|
| 886 |
+
"poetry": {
|
| 887 |
+
"rank": 0.013119533527696792,
|
| 888 |
+
"z": 2.103769786199745,
|
| 889 |
+
"n_pairs": 98
|
| 890 |
+
},
|
| 891 |
+
"typo": {
|
| 892 |
+
"rank": 0.2604166666666667,
|
| 893 |
+
"z": 3.3664332094291844,
|
| 894 |
+
"n_pairs": 96
|
| 895 |
+
}
|
| 896 |
+
},
|
| 897 |
+
"J|Github|s2": {
|
| 898 |
+
"multihop": {
|
| 899 |
+
"rank": 0.3398058252427184,
|
| 900 |
+
"z": 3.787404161055111,
|
| 901 |
+
"n_pairs": 103
|
| 902 |
+
},
|
| 903 |
+
"multilingual": {
|
| 904 |
+
"rank": 0.14176939811457578,
|
| 905 |
+
"z": 2.774661488793223,
|
| 906 |
+
"n_pairs": 394
|
| 907 |
+
},
|
| 908 |
+
"order-ops": {
|
| 909 |
+
"rank": 0.5701298701298702,
|
| 910 |
+
"z": 5.435368727553975,
|
| 911 |
+
"n_pairs": 110
|
| 912 |
+
},
|
| 913 |
+
"poetry": {
|
| 914 |
+
"rank": 0.01603498542274052,
|
| 915 |
+
"z": 2.054967626625178,
|
| 916 |
+
"n_pairs": 98
|
| 917 |
+
},
|
| 918 |
+
"typo": {
|
| 919 |
+
"rank": 0.22470238095238093,
|
| 920 |
+
"z": 3.2467462296287217,
|
| 921 |
+
"n_pairs": 96
|
| 922 |
+
}
|
| 923 |
+
}
|
| 924 |
+
},
|
| 925 |
+
"fits": {
|
| 926 |
+
"J|Wikipedia_en|s0": {
|
| 927 |
+
"n_prompts": 25,
|
| 928 |
+
"fit_seconds": 272.6,
|
| 929 |
+
"prompts_sha256": "9e54a0769bd230f55ba825a256e03c0f34730227e20dfc1b75efc797f711db71",
|
| 930 |
+
"corpus": "Wikipedia_en",
|
| 931 |
+
"seed_block": 0,
|
| 932 |
+
"fro_norm_per_layer": {
|
| 933 |
+
"9": 23.294700622558594,
|
| 934 |
+
"10": 25.415515899658203,
|
| 935 |
+
"11": 28.33647918701172,
|
| 936 |
+
"12": 30.57115936279297,
|
| 937 |
+
"13": 33.302181243896484,
|
| 938 |
+
"14": 36.702049255371094,
|
| 939 |
+
"15": 40.84243392944336,
|
| 940 |
+
"16": 42.165000915527344,
|
| 941 |
+
"17": 44.4361457824707,
|
| 942 |
+
"18": 44.770687103271484,
|
| 943 |
+
"19": 44.682952880859375,
|
| 944 |
+
"20": 45.84259033203125,
|
| 945 |
+
"21": 46.23477554321289
|
| 946 |
+
}
|
| 947 |
+
},
|
| 948 |
+
"J|Wikipedia_en|s1": {
|
| 949 |
+
"n_prompts": 25,
|
| 950 |
+
"fit_seconds": 275.1,
|
| 951 |
+
"prompts_sha256": "2ee5bb179eac0bd1859f75821abcde75138e042cf6c0d93e091acb78931a9fad",
|
| 952 |
+
"corpus": "Wikipedia_en",
|
| 953 |
+
"seed_block": 1,
|
| 954 |
+
"fro_norm_per_layer": {
|
| 955 |
+
"9": 23.690460205078125,
|
| 956 |
+
"10": 25.781587600708008,
|
| 957 |
+
"11": 28.686138153076172,
|
| 958 |
+
"12": 30.9707088470459,
|
| 959 |
+
"13": 33.714210510253906,
|
| 960 |
+
"14": 37.1748161315918,
|
| 961 |
+
"15": 41.22645950317383,
|
| 962 |
+
"16": 42.623661041259766,
|
| 963 |
+
"17": 44.789859771728516,
|
| 964 |
+
"18": 45.092281341552734,
|
| 965 |
+
"19": 44.922061920166016,
|
| 966 |
+
"20": 46.048622131347656,
|
| 967 |
+
"21": 46.331031799316406
|
| 968 |
+
}
|
| 969 |
+
},
|
| 970 |
+
"J|Wikipedia_en|s2": {
|
| 971 |
+
"n_prompts": 25,
|
| 972 |
+
"fit_seconds": 275.9,
|
| 973 |
+
"prompts_sha256": "dcd7a38d916ab5bc0adf8bff4400eb59938b7991e201e082cdc0164d7642659b",
|
| 974 |
+
"corpus": "Wikipedia_en",
|
| 975 |
+
"seed_block": 2,
|
| 976 |
+
"fro_norm_per_layer": {
|
| 977 |
+
"9": 23.420244216918945,
|
| 978 |
+
"10": 25.618276596069336,
|
| 979 |
+
"11": 28.51605796813965,
|
| 980 |
+
"12": 30.880279541015625,
|
| 981 |
+
"13": 33.711265563964844,
|
| 982 |
+
"14": 37.20738983154297,
|
| 983 |
+
"15": 41.2188720703125,
|
| 984 |
+
"16": 42.609657287597656,
|
| 985 |
+
"17": 44.832557678222656,
|
| 986 |
+
"18": 45.15241622924805,
|
| 987 |
+
"19": 44.87080764770508,
|
| 988 |
+
"20": 46.04928970336914,
|
| 989 |
+
"21": 46.32510757446289
|
| 990 |
+
}
|
| 991 |
+
},
|
| 992 |
+
"J|USPTO_Backgrounds|s0": {
|
| 993 |
+
"n_prompts": 25,
|
| 994 |
+
"fit_seconds": 276.3,
|
| 995 |
+
"prompts_sha256": "5a8f7c90631d3924019d98c6c3a93febb4702b0267d73e3bc369a8a1e7b824e4",
|
| 996 |
+
"corpus": "USPTO_Backgrounds",
|
| 997 |
+
"seed_block": 0,
|
| 998 |
+
"fro_norm_per_layer": {
|
| 999 |
+
"9": 27.90863037109375,
|
| 1000 |
+
"10": 30.208759307861328,
|
| 1001 |
+
"11": 32.88587951660156,
|
| 1002 |
+
"12": 35.370208740234375,
|
| 1003 |
+
"13": 38.3754768371582,
|
| 1004 |
+
"14": 41.75584411621094,
|
| 1005 |
+
"15": 45.2714958190918,
|
| 1006 |
+
"16": 46.36663818359375,
|
| 1007 |
+
"17": 47.852264404296875,
|
| 1008 |
+
"18": 47.41462707519531,
|
| 1009 |
+
"19": 46.47177505493164,
|
| 1010 |
+
"20": 47.27696228027344,
|
| 1011 |
+
"21": 46.95705032348633
|
| 1012 |
+
}
|
| 1013 |
+
},
|
| 1014 |
+
"J|USPTO_Backgrounds|s1": {
|
| 1015 |
+
"n_prompts": 25,
|
| 1016 |
+
"fit_seconds": 276.5,
|
| 1017 |
+
"prompts_sha256": "2ec94c2fa4eb9cb721a09d8f1ecd51be4392b6c77ff95f9b66fb173812f04aa3",
|
| 1018 |
+
"corpus": "USPTO_Backgrounds",
|
| 1019 |
+
"seed_block": 1,
|
| 1020 |
+
"fro_norm_per_layer": {
|
| 1021 |
+
"9": 28.49028778076172,
|
| 1022 |
+
"10": 30.822561264038086,
|
| 1023 |
+
"11": 33.30125045776367,
|
| 1024 |
+
"12": 35.712852478027344,
|
| 1025 |
+
"13": 38.649085998535156,
|
| 1026 |
+
"14": 41.987709045410156,
|
| 1027 |
+
"15": 45.43867111206055,
|
| 1028 |
+
"16": 46.37141036987305,
|
| 1029 |
+
"17": 47.787593841552734,
|
| 1030 |
+
"18": 47.34893035888672,
|
| 1031 |
+
"19": 46.38490676879883,
|
| 1032 |
+
"20": 47.20484924316406,
|
| 1033 |
+
"21": 46.929969787597656
|
| 1034 |
+
}
|
| 1035 |
+
},
|
| 1036 |
+
"J|USPTO_Backgrounds|s2": {
|
| 1037 |
+
"n_prompts": 25,
|
| 1038 |
+
"fit_seconds": 276.4,
|
| 1039 |
+
"prompts_sha256": "97174f19e4276737d5949b098c9ee4c669f51f178d9d499b1a5c86b6faf4afad",
|
| 1040 |
+
"corpus": "USPTO_Backgrounds",
|
| 1041 |
+
"seed_block": 2,
|
| 1042 |
+
"fro_norm_per_layer": {
|
| 1043 |
+
"9": 28.800617218017578,
|
| 1044 |
+
"10": 31.102834701538086,
|
| 1045 |
+
"11": 33.716331481933594,
|
| 1046 |
+
"12": 36.08961486816406,
|
| 1047 |
+
"13": 39.015830993652344,
|
| 1048 |
+
"14": 42.280330657958984,
|
| 1049 |
+
"15": 45.75420379638672,
|
| 1050 |
+
"16": 46.61201477050781,
|
| 1051 |
+
"17": 47.96596145629883,
|
| 1052 |
+
"18": 47.46554946899414,
|
| 1053 |
+
"19": 46.47589111328125,
|
| 1054 |
+
"20": 47.297393798828125,
|
| 1055 |
+
"21": 46.97040939331055
|
| 1056 |
+
}
|
| 1057 |
+
},
|
| 1058 |
+
"J|Pile-CC|s0": {
|
| 1059 |
+
"n_prompts": 25,
|
| 1060 |
+
"fit_seconds": 276.6,
|
| 1061 |
+
"prompts_sha256": "e0757ef16e1a8a552d467cf839e49c86eb7e90c7f090efa560e063141bf88c8e",
|
| 1062 |
+
"corpus": "Pile-CC",
|
| 1063 |
+
"seed_block": 0,
|
| 1064 |
+
"fro_norm_per_layer": {
|
| 1065 |
+
"9": 24.955005645751953,
|
| 1066 |
+
"10": 27.270200729370117,
|
| 1067 |
+
"11": 29.907175064086914,
|
| 1068 |
+
"12": 32.539894104003906,
|
| 1069 |
+
"13": 35.55636215209961,
|
| 1070 |
+
"14": 39.129798889160156,
|
| 1071 |
+
"15": 43.26397705078125,
|
| 1072 |
+
"16": 44.49295425415039,
|
| 1073 |
+
"17": 46.42964553833008,
|
| 1074 |
+
"18": 46.149452209472656,
|
| 1075 |
+
"19": 45.46807861328125,
|
| 1076 |
+
"20": 46.5279541015625,
|
| 1077 |
+
"21": 46.48268508911133
|
| 1078 |
+
}
|
| 1079 |
+
},
|
| 1080 |
+
"J|Pile-CC|s1": {
|
| 1081 |
+
"n_prompts": 25,
|
| 1082 |
+
"fit_seconds": 276.7,
|
| 1083 |
+
"prompts_sha256": "3dfe7843f22cef9b3d308d1f1cec3cade456f380b53ba1b8db8776ebe7fed77c",
|
| 1084 |
+
"corpus": "Pile-CC",
|
| 1085 |
+
"seed_block": 1,
|
| 1086 |
+
"fro_norm_per_layer": {
|
| 1087 |
+
"9": 25.38190269470215,
|
| 1088 |
+
"10": 27.569387435913086,
|
| 1089 |
+
"11": 30.241737365722656,
|
| 1090 |
+
"12": 32.736114501953125,
|
| 1091 |
+
"13": 35.672340393066406,
|
| 1092 |
+
"14": 39.08313751220703,
|
| 1093 |
+
"15": 43.28975296020508,
|
| 1094 |
+
"16": 44.42156219482422,
|
| 1095 |
+
"17": 46.440616607666016,
|
| 1096 |
+
"18": 46.22185516357422,
|
| 1097 |
+
"19": 45.58104705810547,
|
| 1098 |
+
"20": 46.608673095703125,
|
| 1099 |
+
"21": 46.55390930175781
|
| 1100 |
+
}
|
| 1101 |
+
},
|
| 1102 |
+
"J|Pile-CC|s2": {
|
| 1103 |
+
"n_prompts": 25,
|
| 1104 |
+
"fit_seconds": 276.6,
|
| 1105 |
+
"prompts_sha256": "782515fa0a56b1398a7237cda29cfac2a4e75c3bf91d1ffba99c86a19d3ab5aa",
|
| 1106 |
+
"corpus": "Pile-CC",
|
| 1107 |
+
"seed_block": 2,
|
| 1108 |
+
"fro_norm_per_layer": {
|
| 1109 |
+
"9": 25.042362213134766,
|
| 1110 |
+
"10": 27.282052993774414,
|
| 1111 |
+
"11": 29.89794158935547,
|
| 1112 |
+
"12": 32.4775505065918,
|
| 1113 |
+
"13": 35.40144348144531,
|
| 1114 |
+
"14": 38.882041931152344,
|
| 1115 |
+
"15": 43.07960510253906,
|
| 1116 |
+
"16": 44.226409912109375,
|
| 1117 |
+
"17": 46.159019470214844,
|
| 1118 |
+
"18": 45.96290588378906,
|
| 1119 |
+
"19": 45.31592559814453,
|
| 1120 |
+
"20": 46.45462417602539,
|
| 1121 |
+
"21": 46.471435546875
|
| 1122 |
+
}
|
| 1123 |
+
},
|
| 1124 |
+
"J|StackExchange|s0": {
|
| 1125 |
+
"n_prompts": 25,
|
| 1126 |
+
"fit_seconds": 277.0,
|
| 1127 |
+
"prompts_sha256": "fb838aed6ad4bb5655506149821e49a7be3de1e40422abf3e744afa0fd80df39",
|
| 1128 |
+
"corpus": "StackExchange",
|
| 1129 |
+
"seed_block": 0,
|
| 1130 |
+
"fro_norm_per_layer": {
|
| 1131 |
+
"9": 26.929075241088867,
|
| 1132 |
+
"10": 29.292173385620117,
|
| 1133 |
+
"11": 31.924766540527344,
|
| 1134 |
+
"12": 34.485313415527344,
|
| 1135 |
+
"13": 37.15853500366211,
|
| 1136 |
+
"14": 40.80641555786133,
|
| 1137 |
+
"15": 44.48529815673828,
|
| 1138 |
+
"16": 45.89080810546875,
|
| 1139 |
+
"17": 47.45210266113281,
|
| 1140 |
+
"18": 47.164772033691406,
|
| 1141 |
+
"19": 46.41997528076172,
|
| 1142 |
+
"20": 47.250492095947266,
|
| 1143 |
+
"21": 46.89768600463867
|
| 1144 |
+
}
|
| 1145 |
+
},
|
| 1146 |
+
"J|StackExchange|s1": {
|
| 1147 |
+
"n_prompts": 25,
|
| 1148 |
+
"fit_seconds": 276.4,
|
| 1149 |
+
"prompts_sha256": "99d6fe9822aacd471eebbf96e115750a6247e2e1bef7353640f9461dfe1aad6b",
|
| 1150 |
+
"corpus": "StackExchange",
|
| 1151 |
+
"seed_block": 1,
|
| 1152 |
+
"fro_norm_per_layer": {
|
| 1153 |
+
"9": 27.617984771728516,
|
| 1154 |
+
"10": 30.13306427001953,
|
| 1155 |
+
"11": 32.612632751464844,
|
| 1156 |
+
"12": 35.15741729736328,
|
| 1157 |
+
"13": 37.79157257080078,
|
| 1158 |
+
"14": 41.37997055053711,
|
| 1159 |
+
"15": 45.01513671875,
|
| 1160 |
+
"16": 46.328369140625,
|
| 1161 |
+
"17": 47.70050811767578,
|
| 1162 |
+
"18": 47.40119552612305,
|
| 1163 |
+
"19": 46.614158630371094,
|
| 1164 |
+
"20": 47.42560958862305,
|
| 1165 |
+
"21": 47.04204559326172
|
| 1166 |
+
}
|
| 1167 |
+
},
|
| 1168 |
+
"J|StackExchange|s2": {
|
| 1169 |
+
"n_prompts": 25,
|
| 1170 |
+
"fit_seconds": 276.3,
|
| 1171 |
+
"prompts_sha256": "094be840f776043c464c6dc2ada90ee60b7a6349a4faf5edbee6f7646e048e58",
|
| 1172 |
+
"corpus": "StackExchange",
|
| 1173 |
+
"seed_block": 2,
|
| 1174 |
+
"fro_norm_per_layer": {
|
| 1175 |
+
"9": 27.36570930480957,
|
| 1176 |
+
"10": 29.822795867919922,
|
| 1177 |
+
"11": 32.43924331665039,
|
| 1178 |
+
"12": 35.0574836730957,
|
| 1179 |
+
"13": 37.744300842285156,
|
| 1180 |
+
"14": 41.3569221496582,
|
| 1181 |
+
"15": 45.00175476074219,
|
| 1182 |
+
"16": 46.438682556152344,
|
| 1183 |
+
"17": 47.874332427978516,
|
| 1184 |
+
"18": 47.57386779785156,
|
| 1185 |
+
"19": 46.77056884765625,
|
| 1186 |
+
"20": 47.53171157836914,
|
| 1187 |
+
"21": 47.10435104370117
|
| 1188 |
+
}
|
| 1189 |
+
},
|
| 1190 |
+
"J|Github|s0": {
|
| 1191 |
+
"n_prompts": 25,
|
| 1192 |
+
"fit_seconds": 275.9,
|
| 1193 |
+
"prompts_sha256": "fb6856c864ad3d738b939e1ca21b65a129f645bf18dd0221c4c9e2c1dadb7fe5",
|
| 1194 |
+
"corpus": "Github",
|
| 1195 |
+
"seed_block": 0,
|
| 1196 |
+
"fro_norm_per_layer": {
|
| 1197 |
+
"9": 23.499574661254883,
|
| 1198 |
+
"10": 25.101781845092773,
|
| 1199 |
+
"11": 27.236875534057617,
|
| 1200 |
+
"12": 29.104747772216797,
|
| 1201 |
+
"13": 31.28536033630371,
|
| 1202 |
+
"14": 34.06266403198242,
|
| 1203 |
+
"15": 37.80937957763672,
|
| 1204 |
+
"16": 39.80789566040039,
|
| 1205 |
+
"17": 42.34746551513672,
|
| 1206 |
+
"18": 43.70899200439453,
|
| 1207 |
+
"19": 44.678932189941406,
|
| 1208 |
+
"20": 45.65796661376953,
|
| 1209 |
+
"21": 46.37763214111328
|
| 1210 |
+
}
|
| 1211 |
+
},
|
| 1212 |
+
"J|Github|s1": {
|
| 1213 |
+
"n_prompts": 25,
|
| 1214 |
+
"fit_seconds": 275.3,
|
| 1215 |
+
"prompts_sha256": "8acd4ae02d98b1cb29ad26750a7f98b3f9bb842226f217fbc2ab18547092f0a1",
|
| 1216 |
+
"corpus": "Github",
|
| 1217 |
+
"seed_block": 1,
|
| 1218 |
+
"fro_norm_per_layer": {
|
| 1219 |
+
"9": 24.057697296142578,
|
| 1220 |
+
"10": 25.858362197875977,
|
| 1221 |
+
"11": 28.058307647705078,
|
| 1222 |
+
"12": 30.10573387145996,
|
| 1223 |
+
"13": 32.42531204223633,
|
| 1224 |
+
"14": 35.27041244506836,
|
| 1225 |
+
"15": 39.16661071777344,
|
| 1226 |
+
"16": 41.07899856567383,
|
| 1227 |
+
"17": 43.373783111572266,
|
| 1228 |
+
"18": 44.43169403076172,
|
| 1229 |
+
"19": 45.091068267822266,
|
| 1230 |
+
"20": 46.02471160888672,
|
| 1231 |
+
"21": 46.53116226196289
|
| 1232 |
+
}
|
| 1233 |
+
},
|
| 1234 |
+
"J|Github|s2": {
|
| 1235 |
+
"n_prompts": 25,
|
| 1236 |
+
"fit_seconds": 275.3,
|
| 1237 |
+
"prompts_sha256": "fb8e687f84bbaca1821949876dabd6beb24389556d54c1fa6d9234184037fc68",
|
| 1238 |
+
"corpus": "Github",
|
| 1239 |
+
"seed_block": 2,
|
| 1240 |
+
"fro_norm_per_layer": {
|
| 1241 |
+
"9": 23.682188034057617,
|
| 1242 |
+
"10": 25.320621490478516,
|
| 1243 |
+
"11": 27.47364044189453,
|
| 1244 |
+
"12": 29.382030487060547,
|
| 1245 |
+
"13": 31.588592529296875,
|
| 1246 |
+
"14": 34.38534927368164,
|
| 1247 |
+
"15": 38.232765197753906,
|
| 1248 |
+
"16": 40.23326873779297,
|
| 1249 |
+
"17": 42.76146697998047,
|
| 1250 |
+
"18": 44.06857681274414,
|
| 1251 |
+
"19": 44.8614501953125,
|
| 1252 |
+
"20": 45.87016677856445,
|
| 1253 |
+
"21": 46.53526306152344
|
| 1254 |
+
}
|
| 1255 |
+
}
|
| 1256 |
+
},
|
| 1257 |
+
"logit_I_pooled_rank_flatmean7": 0.23942414367877324,
|
| 1258 |
+
"provenance": {
|
| 1259 |
+
"schema_version": 1,
|
| 1260 |
+
"experiment": "CV6",
|
| 1261 |
+
"utc": "2026-08-24T00:26:12+00:00",
|
| 1262 |
+
"script": {
|
| 1263 |
+
"path": "experiments/cv6_per_family_ladder.py",
|
| 1264 |
+
"sha256": "f416c3cb55e95a02315504066479c2f2ba073c0d4676bbf09adb2f481cdb4a5b",
|
| 1265 |
+
"n_lines": 750
|
| 1266 |
+
},
|
| 1267 |
+
"git": {
|
| 1268 |
+
"commit": null,
|
| 1269 |
+
"commit_short": null,
|
| 1270 |
+
"branch": null,
|
| 1271 |
+
"dirty": false,
|
| 1272 |
+
"dirty_files": [],
|
| 1273 |
+
"dirty_diff_sha256": null,
|
| 1274 |
+
"describe": null
|
| 1275 |
+
},
|
| 1276 |
+
"argv": [
|
| 1277 |
+
"experiments/cv6_per_family_ladder.py",
|
| 1278 |
+
"--model",
|
| 1279 |
+
"1.4b",
|
| 1280 |
+
"--device",
|
| 1281 |
+
"cuda",
|
| 1282 |
+
"--cache-out",
|
| 1283 |
+
"results/cv6/ht_cache_1.4b.pt",
|
| 1284 |
+
"--save-lens-dir",
|
| 1285 |
+
"results/cv6/lenses"
|
| 1286 |
+
],
|
| 1287 |
+
"cwd": ".",
|
| 1288 |
+
"inputs": [
|
| 1289 |
+
{
|
| 1290 |
+
"path": "corpora/Wikipedia_en.jsonl",
|
| 1291 |
+
"sha256": "b30f4cef11add560f7df8ed07a613e3fcc6a4dda865975bec15f2952253bda77",
|
| 1292 |
+
"bytes": 10329196,
|
| 1293 |
+
"exists": true
|
| 1294 |
+
},
|
| 1295 |
+
{
|
| 1296 |
+
"path": "corpora/USPTO_Backgrounds.jsonl",
|
| 1297 |
+
"sha256": "b09ac74bdc8b169ee5c9bea6be481262f743947a92107164f30eca2346e3ccb9",
|
| 1298 |
+
"bytes": 10221054,
|
| 1299 |
+
"exists": true
|
| 1300 |
+
},
|
| 1301 |
+
{
|
| 1302 |
+
"path": "corpora/Pile-CC.jsonl",
|
| 1303 |
+
"sha256": "abbb9f19a6eb6f4c12fb382735832ceec271fece15cfebcf0fe838765d4b1e74",
|
| 1304 |
+
"bytes": 13512530,
|
| 1305 |
+
"exists": true
|
| 1306 |
+
},
|
| 1307 |
+
{
|
| 1308 |
+
"path": "corpora/StackExchange.jsonl",
|
| 1309 |
+
"sha256": "bc25e2d5775e9719ebcbdaa3bd115d827f72ebf11f47fedb377f8b0a69a28971",
|
| 1310 |
+
"bytes": 5980156,
|
| 1311 |
+
"exists": true
|
| 1312 |
+
},
|
| 1313 |
+
{
|
| 1314 |
+
"path": "corpora/Github.jsonl",
|
| 1315 |
+
"sha256": "f2bb459d10cad6c13a8b66c86191bc43df4393ba5b722a5315ea878fad2d1a26",
|
| 1316 |
+
"bytes": 17486378,
|
| 1317 |
+
"exists": true
|
| 1318 |
+
}
|
| 1319 |
+
],
|
| 1320 |
+
"env": {
|
| 1321 |
+
"python": "3.11.9",
|
| 1322 |
+
"torch": "2.11.0+cu128",
|
| 1323 |
+
"transformers": "5.15.1",
|
| 1324 |
+
"numpy": "2.4.6",
|
| 1325 |
+
"platform": "Linux-5.15.0-157-generic-x86_64-with-glibc2.35",
|
| 1326 |
+
"machine": "x86_64",
|
| 1327 |
+
"hostname": "<redacted for double-blind review>",
|
| 1328 |
+
"tf32": {
|
| 1329 |
+
"allow_tf32_matmul": false,
|
| 1330 |
+
"allow_tf32_cudnn": false,
|
| 1331 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 1332 |
+
}
|
| 1333 |
+
},
|
| 1334 |
+
"payload_sha256": "307871427cdc9dbafbe962115e903a649c5f232dc75ac91163e4dd3832b28687"
|
| 1335 |
+
}
|
| 1336 |
+
}
|
results/cv6/cv6_2.8b_n25.json
ADDED
|
@@ -0,0 +1,1416 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV6 \u2014 the per-family corpus contrast, one model",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 4 |
+
"status": "PRE-REGISTERED",
|
| 5 |
+
"decision_rule_verbatim": "Evaluated PER FAMILY; the verdict is the count across the five. REPLICATES if R(f,2.8B) >= 10 on >= 4 of 5 families AND ordering tau >= 0.6 on >= 3 of 5. ATTENUATES if R(f,2.8B) < 5 on >= 4 of 5 families. UNCLEAR otherwise, including a split where some families replicate and others do not \u2014 a split is an informative outcome, not a failure. Do not pool to force a verdict. Do not add models to chase a threshold. Do not drop a family.",
|
| 6 |
+
"model": "EleutherAI/pythia-2.8b-deduped",
|
| 7 |
+
"short": "2.8b",
|
| 8 |
+
"n_layers": 32,
|
| 9 |
+
"d_model": 2560,
|
| 10 |
+
"band": [
|
| 11 |
+
12,
|
| 12 |
+
13,
|
| 13 |
+
14,
|
| 14 |
+
15,
|
| 15 |
+
16,
|
| 16 |
+
17,
|
| 17 |
+
18,
|
| 18 |
+
19,
|
| 19 |
+
20,
|
| 20 |
+
21,
|
| 21 |
+
22,
|
| 22 |
+
23,
|
| 23 |
+
24,
|
| 24 |
+
25,
|
| 25 |
+
26,
|
| 26 |
+
27,
|
| 27 |
+
28,
|
| 28 |
+
29
|
| 29 |
+
],
|
| 30 |
+
"band_rule": "floor(0.38*L)..floor(0.92*L), layers < L-2",
|
| 31 |
+
"band_matches_rule": true,
|
| 32 |
+
"N": 25,
|
| 33 |
+
"dim_batch": 128,
|
| 34 |
+
"max_seq_len": 128,
|
| 35 |
+
"skip_first": 16,
|
| 36 |
+
"target_layer_arg": -2,
|
| 37 |
+
"target_layer_effective": 30,
|
| 38 |
+
"device": "cuda",
|
| 39 |
+
"dtype": "float32",
|
| 40 |
+
"tf32": {
|
| 41 |
+
"matmul_allow_tf32": false,
|
| 42 |
+
"cudnn_allow_tf32": false,
|
| 43 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 44 |
+
},
|
| 45 |
+
"gpu": "NVIDIA H100 NVL",
|
| 46 |
+
"gpu_total_gb": 93.08551025390625,
|
| 47 |
+
"peak_vram_gb": 46.3552303314209,
|
| 48 |
+
"torch": "2.11.0+cu128",
|
| 49 |
+
"git_commit_at_launch": "518d192e9e215603957c85ac4a1308153bf4cf06",
|
| 50 |
+
"K": [
|
| 51 |
+
1,
|
| 52 |
+
2,
|
| 53 |
+
5,
|
| 54 |
+
10,
|
| 55 |
+
20,
|
| 56 |
+
50,
|
| 57 |
+
100
|
| 58 |
+
],
|
| 59 |
+
"admitted_sets": [
|
| 60 |
+
"multihop",
|
| 61 |
+
"multilingual",
|
| 62 |
+
"order-ops",
|
| 63 |
+
"poetry",
|
| 64 |
+
"typo"
|
| 65 |
+
],
|
| 66 |
+
"corpora": [
|
| 67 |
+
"Wikipedia_en",
|
| 68 |
+
"USPTO_Backgrounds",
|
| 69 |
+
"Pile-CC",
|
| 70 |
+
"StackExchange",
|
| 71 |
+
"Github"
|
| 72 |
+
],
|
| 73 |
+
"seeds": [
|
| 74 |
+
0,
|
| 75 |
+
1,
|
| 76 |
+
2
|
| 77 |
+
],
|
| 78 |
+
"readout_convention": "STRIPPED (corrected), no prefix, flat-mean-7",
|
| 79 |
+
"pooling_convention": "flat pool over (item, intermediate) pairs WITHIN a family; rank = min over band then flat-mean-7 pass@k; z = max over band then mean over pairs",
|
| 80 |
+
"n_items": 449,
|
| 81 |
+
"n_pairs": 801,
|
| 82 |
+
"n_pairs_per_family": {
|
| 83 |
+
"multihop": 103,
|
| 84 |
+
"multilingual": 394,
|
| 85 |
+
"order-ops": 110,
|
| 86 |
+
"poetry": 98,
|
| 87 |
+
"typo": 96
|
| 88 |
+
},
|
| 89 |
+
"per_arm": {
|
| 90 |
+
"logit_I": {
|
| 91 |
+
"multihop": {
|
| 92 |
+
"rank": 0.27600554785020803,
|
| 93 |
+
"z": 3.693257671826094,
|
| 94 |
+
"n_pairs": 103
|
| 95 |
+
},
|
| 96 |
+
"multilingual": {
|
| 97 |
+
"rank": 0.22588832487309643,
|
| 98 |
+
"z": 3.0744364978603755,
|
| 99 |
+
"n_pairs": 394
|
| 100 |
+
},
|
| 101 |
+
"order-ops": {
|
| 102 |
+
"rank": 0.538961038961039,
|
| 103 |
+
"z": 5.5365345900708975,
|
| 104 |
+
"n_pairs": 110
|
| 105 |
+
},
|
| 106 |
+
"poetry": {
|
| 107 |
+
"rank": 0.02769679300291545,
|
| 108 |
+
"z": 2.430429516398177,
|
| 109 |
+
"n_pairs": 98
|
| 110 |
+
},
|
| 111 |
+
"typo": {
|
| 112 |
+
"rank": 0.28125,
|
| 113 |
+
"z": 3.6758319437503815,
|
| 114 |
+
"n_pairs": 96
|
| 115 |
+
}
|
| 116 |
+
},
|
| 117 |
+
"J|Wikipedia_en|s0": {
|
| 118 |
+
"multihop": {
|
| 119 |
+
"rank": 0.37170596393897365,
|
| 120 |
+
"z": 3.748083519704134,
|
| 121 |
+
"n_pairs": 103
|
| 122 |
+
},
|
| 123 |
+
"multilingual": {
|
| 124 |
+
"rank": 0.3176214648295867,
|
| 125 |
+
"z": 3.4327350325390773,
|
| 126 |
+
"n_pairs": 394
|
| 127 |
+
},
|
| 128 |
+
"order-ops": {
|
| 129 |
+
"rank": 0.5688311688311688,
|
| 130 |
+
"z": 5.169967714222994,
|
| 131 |
+
"n_pairs": 110
|
| 132 |
+
},
|
| 133 |
+
"poetry": {
|
| 134 |
+
"rank": 0.06268221574344024,
|
| 135 |
+
"z": 2.520261741414362,
|
| 136 |
+
"n_pairs": 98
|
| 137 |
+
},
|
| 138 |
+
"typo": {
|
| 139 |
+
"rank": 0.2976190476190476,
|
| 140 |
+
"z": 3.459335678567489,
|
| 141 |
+
"n_pairs": 96
|
| 142 |
+
}
|
| 143 |
+
},
|
| 144 |
+
"J|Wikipedia_en|s1": {
|
| 145 |
+
"multihop": {
|
| 146 |
+
"rank": 0.37586685159500693,
|
| 147 |
+
"z": 3.7837847610121793,
|
| 148 |
+
"n_pairs": 103
|
| 149 |
+
},
|
| 150 |
+
"multilingual": {
|
| 151 |
+
"rank": 0.31725888324873097,
|
| 152 |
+
"z": 3.460326204445156,
|
| 153 |
+
"n_pairs": 394
|
| 154 |
+
},
|
| 155 |
+
"order-ops": {
|
| 156 |
+
"rank": 0.5844155844155845,
|
| 157 |
+
"z": 5.257566994970495,
|
| 158 |
+
"n_pairs": 110
|
| 159 |
+
},
|
| 160 |
+
"poetry": {
|
| 161 |
+
"rank": 0.06268221574344023,
|
| 162 |
+
"z": 2.5308002853880125,
|
| 163 |
+
"n_pairs": 98
|
| 164 |
+
},
|
| 165 |
+
"typo": {
|
| 166 |
+
"rank": 0.3199404761904762,
|
| 167 |
+
"z": 3.5066992342472076,
|
| 168 |
+
"n_pairs": 96
|
| 169 |
+
}
|
| 170 |
+
},
|
| 171 |
+
"J|Wikipedia_en|s2": {
|
| 172 |
+
"multihop": {
|
| 173 |
+
"rank": 0.36893203883495146,
|
| 174 |
+
"z": 3.7126920130646344,
|
| 175 |
+
"n_pairs": 103
|
| 176 |
+
},
|
| 177 |
+
"multilingual": {
|
| 178 |
+
"rank": 0.3110949963741842,
|
| 179 |
+
"z": 3.4116927991966306,
|
| 180 |
+
"n_pairs": 394
|
| 181 |
+
},
|
| 182 |
+
"order-ops": {
|
| 183 |
+
"rank": 0.5805194805194805,
|
| 184 |
+
"z": 5.221894264221191,
|
| 185 |
+
"n_pairs": 110
|
| 186 |
+
},
|
| 187 |
+
"poetry": {
|
| 188 |
+
"rank": 0.06268221574344023,
|
| 189 |
+
"z": 2.5136427721198724,
|
| 190 |
+
"n_pairs": 98
|
| 191 |
+
},
|
| 192 |
+
"typo": {
|
| 193 |
+
"rank": 0.3095238095238096,
|
| 194 |
+
"z": 3.4875277963777385,
|
| 195 |
+
"n_pairs": 96
|
| 196 |
+
}
|
| 197 |
+
},
|
| 198 |
+
"J|USPTO_Backgrounds|s0": {
|
| 199 |
+
"multihop": {
|
| 200 |
+
"rank": 0.420249653259362,
|
| 201 |
+
"z": 3.9431607584351474,
|
| 202 |
+
"n_pairs": 103
|
| 203 |
+
},
|
| 204 |
+
"multilingual": {
|
| 205 |
+
"rank": 0.3328498912255258,
|
| 206 |
+
"z": 3.617218898637646,
|
| 207 |
+
"n_pairs": 394
|
| 208 |
+
},
|
| 209 |
+
"order-ops": {
|
| 210 |
+
"rank": 0.6246753246753247,
|
| 211 |
+
"z": 5.513402173735878,
|
| 212 |
+
"n_pairs": 110
|
| 213 |
+
},
|
| 214 |
+
"poetry": {
|
| 215 |
+
"rank": 0.08454810495626822,
|
| 216 |
+
"z": 2.645613257982293,
|
| 217 |
+
"n_pairs": 98
|
| 218 |
+
},
|
| 219 |
+
"typo": {
|
| 220 |
+
"rank": 0.42708333333333337,
|
| 221 |
+
"z": 3.7516257812579474,
|
| 222 |
+
"n_pairs": 96
|
| 223 |
+
}
|
| 224 |
+
},
|
| 225 |
+
"J|USPTO_Backgrounds|s1": {
|
| 226 |
+
"multihop": {
|
| 227 |
+
"rank": 0.4133148404993065,
|
| 228 |
+
"z": 3.9598770234191303,
|
| 229 |
+
"n_pairs": 103
|
| 230 |
+
},
|
| 231 |
+
"multilingual": {
|
| 232 |
+
"rank": 0.3390137781000725,
|
| 233 |
+
"z": 3.662088529410096,
|
| 234 |
+
"n_pairs": 394
|
| 235 |
+
},
|
| 236 |
+
"order-ops": {
|
| 237 |
+
"rank": 0.6571428571428571,
|
| 238 |
+
"z": 5.61998973976482,
|
| 239 |
+
"n_pairs": 110
|
| 240 |
+
},
|
| 241 |
+
"poetry": {
|
| 242 |
+
"rank": 0.09329446064139943,
|
| 243 |
+
"z": 2.6892670818737576,
|
| 244 |
+
"n_pairs": 98
|
| 245 |
+
},
|
| 246 |
+
"typo": {
|
| 247 |
+
"rank": 0.44345238095238093,
|
| 248 |
+
"z": 3.785833690315485,
|
| 249 |
+
"n_pairs": 96
|
| 250 |
+
}
|
| 251 |
+
},
|
| 252 |
+
"J|USPTO_Backgrounds|s2": {
|
| 253 |
+
"multihop": {
|
| 254 |
+
"rank": 0.39944521497919555,
|
| 255 |
+
"z": 3.9532608187314375,
|
| 256 |
+
"n_pairs": 103
|
| 257 |
+
},
|
| 258 |
+
"multilingual": {
|
| 259 |
+
"rank": 0.34590282813633066,
|
| 260 |
+
"z": 3.7125451938452456,
|
| 261 |
+
"n_pairs": 394
|
| 262 |
+
},
|
| 263 |
+
"order-ops": {
|
| 264 |
+
"rank": 0.6376623376623376,
|
| 265 |
+
"z": 5.639909208904613,
|
| 266 |
+
"n_pairs": 110
|
| 267 |
+
},
|
| 268 |
+
"poetry": {
|
| 269 |
+
"rank": 0.10787172011661807,
|
| 270 |
+
"z": 2.747300364533249,
|
| 271 |
+
"n_pairs": 98
|
| 272 |
+
},
|
| 273 |
+
"typo": {
|
| 274 |
+
"rank": 0.44642857142857145,
|
| 275 |
+
"z": 3.8546179458498955,
|
| 276 |
+
"n_pairs": 96
|
| 277 |
+
}
|
| 278 |
+
},
|
| 279 |
+
"J|Pile-CC|s0": {
|
| 280 |
+
"multihop": {
|
| 281 |
+
"rank": 0.3828016643550624,
|
| 282 |
+
"z": 3.9039036672092178,
|
| 283 |
+
"n_pairs": 103
|
| 284 |
+
},
|
| 285 |
+
"multilingual": {
|
| 286 |
+
"rank": 0.3332124728063815,
|
| 287 |
+
"z": 3.6563512620889598,
|
| 288 |
+
"n_pairs": 394
|
| 289 |
+
},
|
| 290 |
+
"order-ops": {
|
| 291 |
+
"rank": 0.5922077922077922,
|
| 292 |
+
"z": 5.404743753780018,
|
| 293 |
+
"n_pairs": 110
|
| 294 |
+
},
|
| 295 |
+
"poetry": {
|
| 296 |
+
"rank": 0.06851311953352769,
|
| 297 |
+
"z": 2.5789867603048986,
|
| 298 |
+
"n_pairs": 98
|
| 299 |
+
},
|
| 300 |
+
"typo": {
|
| 301 |
+
"rank": 0.3630952380952381,
|
| 302 |
+
"z": 3.8045580573379993,
|
| 303 |
+
"n_pairs": 96
|
| 304 |
+
}
|
| 305 |
+
},
|
| 306 |
+
"J|Pile-CC|s1": {
|
| 307 |
+
"multihop": {
|
| 308 |
+
"rank": 0.3883495145631068,
|
| 309 |
+
"z": 3.8896664404174657,
|
| 310 |
+
"n_pairs": 103
|
| 311 |
+
},
|
| 312 |
+
"multilingual": {
|
| 313 |
+
"rank": 0.3234227701232778,
|
| 314 |
+
"z": 3.567083712913058,
|
| 315 |
+
"n_pairs": 394
|
| 316 |
+
},
|
| 317 |
+
"order-ops": {
|
| 318 |
+
"rank": 0.5818181818181819,
|
| 319 |
+
"z": 5.46329207257791,
|
| 320 |
+
"n_pairs": 110
|
| 321 |
+
},
|
| 322 |
+
"poetry": {
|
| 323 |
+
"rank": 0.05393586005830903,
|
| 324 |
+
"z": 2.606388902177616,
|
| 325 |
+
"n_pairs": 98
|
| 326 |
+
},
|
| 327 |
+
"typo": {
|
| 328 |
+
"rank": 0.40476190476190477,
|
| 329 |
+
"z": 3.885098062455654,
|
| 330 |
+
"n_pairs": 96
|
| 331 |
+
}
|
| 332 |
+
},
|
| 333 |
+
"J|Pile-CC|s2": {
|
| 334 |
+
"multihop": {
|
| 335 |
+
"rank": 0.391123439667129,
|
| 336 |
+
"z": 3.9153419918226966,
|
| 337 |
+
"n_pairs": 103
|
| 338 |
+
},
|
| 339 |
+
"multilingual": {
|
| 340 |
+
"rank": 0.33575054387237124,
|
| 341 |
+
"z": 3.6062580548264656,
|
| 342 |
+
"n_pairs": 394
|
| 343 |
+
},
|
| 344 |
+
"order-ops": {
|
| 345 |
+
"rank": 0.5909090909090909,
|
| 346 |
+
"z": 5.418533691492947,
|
| 347 |
+
"n_pairs": 110
|
| 348 |
+
},
|
| 349 |
+
"poetry": {
|
| 350 |
+
"rank": 0.06705539358600583,
|
| 351 |
+
"z": 2.5475755419049944,
|
| 352 |
+
"n_pairs": 98
|
| 353 |
+
},
|
| 354 |
+
"typo": {
|
| 355 |
+
"rank": 0.38244047619047616,
|
| 356 |
+
"z": 3.770227551460266,
|
| 357 |
+
"n_pairs": 96
|
| 358 |
+
}
|
| 359 |
+
},
|
| 360 |
+
"J|StackExchange|s0": {
|
| 361 |
+
"multihop": {
|
| 362 |
+
"rank": 0.42718446601941745,
|
| 363 |
+
"z": 4.062357205789066,
|
| 364 |
+
"n_pairs": 103
|
| 365 |
+
},
|
| 366 |
+
"multilingual": {
|
| 367 |
+
"rank": 0.33575054387237124,
|
| 368 |
+
"z": 3.7145613397135953,
|
| 369 |
+
"n_pairs": 394
|
| 370 |
+
},
|
| 371 |
+
"order-ops": {
|
| 372 |
+
"rank": 0.5675324675324674,
|
| 373 |
+
"z": 5.500116468559612,
|
| 374 |
+
"n_pairs": 110
|
| 375 |
+
},
|
| 376 |
+
"poetry": {
|
| 377 |
+
"rank": 0.05830903790087464,
|
| 378 |
+
"z": 2.4863040441153,
|
| 379 |
+
"n_pairs": 98
|
| 380 |
+
},
|
| 381 |
+
"typo": {
|
| 382 |
+
"rank": 0.40029761904761907,
|
| 383 |
+
"z": 3.8565272353589535,
|
| 384 |
+
"n_pairs": 96
|
| 385 |
+
}
|
| 386 |
+
},
|
| 387 |
+
"J|StackExchange|s1": {
|
| 388 |
+
"multihop": {
|
| 389 |
+
"rank": 0.40083217753120665,
|
| 390 |
+
"z": 3.945768759088609,
|
| 391 |
+
"n_pairs": 103
|
| 392 |
+
},
|
| 393 |
+
"multilingual": {
|
| 394 |
+
"rank": 0.34699057287889773,
|
| 395 |
+
"z": 3.717590422497183,
|
| 396 |
+
"n_pairs": 394
|
| 397 |
+
},
|
| 398 |
+
"order-ops": {
|
| 399 |
+
"rank": 0.5597402597402598,
|
| 400 |
+
"z": 5.4544677734375,
|
| 401 |
+
"n_pairs": 110
|
| 402 |
+
},
|
| 403 |
+
"poetry": {
|
| 404 |
+
"rank": 0.05830903790087464,
|
| 405 |
+
"z": 2.4743533736589005,
|
| 406 |
+
"n_pairs": 98
|
| 407 |
+
},
|
| 408 |
+
"typo": {
|
| 409 |
+
"rank": 0.40625,
|
| 410 |
+
"z": 3.858109233280023,
|
| 411 |
+
"n_pairs": 96
|
| 412 |
+
}
|
| 413 |
+
},
|
| 414 |
+
"J|StackExchange|s2": {
|
| 415 |
+
"multihop": {
|
| 416 |
+
"rank": 0.42857142857142866,
|
| 417 |
+
"z": 4.065541328735722,
|
| 418 |
+
"n_pairs": 103
|
| 419 |
+
},
|
| 420 |
+
"multilingual": {
|
| 421 |
+
"rank": 0.3328498912255258,
|
| 422 |
+
"z": 3.696884382043393,
|
| 423 |
+
"n_pairs": 394
|
| 424 |
+
},
|
| 425 |
+
"order-ops": {
|
| 426 |
+
"rank": 0.5857142857142856,
|
| 427 |
+
"z": 5.543083320964467,
|
| 428 |
+
"n_pairs": 110
|
| 429 |
+
},
|
| 430 |
+
"poetry": {
|
| 431 |
+
"rank": 0.05393586005830903,
|
| 432 |
+
"z": 2.4796542653015683,
|
| 433 |
+
"n_pairs": 98
|
| 434 |
+
},
|
| 435 |
+
"typo": {
|
| 436 |
+
"rank": 0.4032738095238095,
|
| 437 |
+
"z": 3.8242265147467456,
|
| 438 |
+
"n_pairs": 96
|
| 439 |
+
}
|
| 440 |
+
},
|
| 441 |
+
"J|Github|s0": {
|
| 442 |
+
"multihop": {
|
| 443 |
+
"rank": 0.32454923717059636,
|
| 444 |
+
"z": 3.357145009688961,
|
| 445 |
+
"n_pairs": 103
|
| 446 |
+
},
|
| 447 |
+
"multilingual": {
|
| 448 |
+
"rank": 0.23567802755620013,
|
| 449 |
+
"z": 2.9966616588195567,
|
| 450 |
+
"n_pairs": 394
|
| 451 |
+
},
|
| 452 |
+
"order-ops": {
|
| 453 |
+
"rank": 0.5999999999999999,
|
| 454 |
+
"z": 5.133839065378362,
|
| 455 |
+
"n_pairs": 110
|
| 456 |
+
},
|
| 457 |
+
"poetry": {
|
| 458 |
+
"rank": 0.030612244897959186,
|
| 459 |
+
"z": 2.274473075964013,
|
| 460 |
+
"n_pairs": 98
|
| 461 |
+
},
|
| 462 |
+
"typo": {
|
| 463 |
+
"rank": 0.2202380952380952,
|
| 464 |
+
"z": 3.1765695475041866,
|
| 465 |
+
"n_pairs": 96
|
| 466 |
+
}
|
| 467 |
+
},
|
| 468 |
+
"J|Github|s1": {
|
| 469 |
+
"multihop": {
|
| 470 |
+
"rank": 0.348127600554785,
|
| 471 |
+
"z": 3.5227397807593483,
|
| 472 |
+
"n_pairs": 103
|
| 473 |
+
},
|
| 474 |
+
"multilingual": {
|
| 475 |
+
"rank": 0.2646845540246555,
|
| 476 |
+
"z": 3.135737712764498,
|
| 477 |
+
"n_pairs": 394
|
| 478 |
+
},
|
| 479 |
+
"order-ops": {
|
| 480 |
+
"rank": 0.5896103896103896,
|
| 481 |
+
"z": 5.305073720758611,
|
| 482 |
+
"n_pairs": 110
|
| 483 |
+
},
|
| 484 |
+
"poetry": {
|
| 485 |
+
"rank": 0.039358600583090375,
|
| 486 |
+
"z": 2.350923566185698,
|
| 487 |
+
"n_pairs": 98
|
| 488 |
+
},
|
| 489 |
+
"typo": {
|
| 490 |
+
"rank": 0.27232142857142855,
|
| 491 |
+
"z": 3.336014381299416,
|
| 492 |
+
"n_pairs": 96
|
| 493 |
+
}
|
| 494 |
+
},
|
| 495 |
+
"J|Github|s2": {
|
| 496 |
+
"multihop": {
|
| 497 |
+
"rank": 0.32316227461858527,
|
| 498 |
+
"z": 3.391732178845452,
|
| 499 |
+
"n_pairs": 103
|
| 500 |
+
},
|
| 501 |
+
"multilingual": {
|
| 502 |
+
"rank": 0.2548948513415518,
|
| 503 |
+
"z": 3.034613796599625,
|
| 504 |
+
"n_pairs": 394
|
| 505 |
+
},
|
| 506 |
+
"order-ops": {
|
| 507 |
+
"rank": 0.5935064935064934,
|
| 508 |
+
"z": 5.225355126641014,
|
| 509 |
+
"n_pairs": 110
|
| 510 |
+
},
|
| 511 |
+
"poetry": {
|
| 512 |
+
"rank": 0.03498542274052478,
|
| 513 |
+
"z": 2.3192731969210567,
|
| 514 |
+
"n_pairs": 98
|
| 515 |
+
},
|
| 516 |
+
"typo": {
|
| 517 |
+
"rank": 0.25595238095238093,
|
| 518 |
+
"z": 3.23986042663455,
|
| 519 |
+
"n_pairs": 96
|
| 520 |
+
}
|
| 521 |
+
}
|
| 522 |
+
},
|
| 523 |
+
"per_arm_fp16_roundtrip": {
|
| 524 |
+
"J|Wikipedia_en|s0": {
|
| 525 |
+
"multihop": {
|
| 526 |
+
"rank": 0.37170596393897365,
|
| 527 |
+
"z": 3.7479759431579738,
|
| 528 |
+
"n_pairs": 103
|
| 529 |
+
},
|
| 530 |
+
"multilingual": {
|
| 531 |
+
"rank": 0.3176214648295867,
|
| 532 |
+
"z": 3.4326538239336255,
|
| 533 |
+
"n_pairs": 394
|
| 534 |
+
},
|
| 535 |
+
"order-ops": {
|
| 536 |
+
"rank": 0.5688311688311688,
|
| 537 |
+
"z": 5.170113795453852,
|
| 538 |
+
"n_pairs": 110
|
| 539 |
+
},
|
| 540 |
+
"poetry": {
|
| 541 |
+
"rank": 0.06268221574344024,
|
| 542 |
+
"z": 2.5202506519093806,
|
| 543 |
+
"n_pairs": 98
|
| 544 |
+
},
|
| 545 |
+
"typo": {
|
| 546 |
+
"rank": 0.2976190476190476,
|
| 547 |
+
"z": 3.459295087804397,
|
| 548 |
+
"n_pairs": 96
|
| 549 |
+
}
|
| 550 |
+
},
|
| 551 |
+
"J|Wikipedia_en|s1": {
|
| 552 |
+
"multihop": {
|
| 553 |
+
"rank": 0.37586685159500693,
|
| 554 |
+
"z": 3.78365947783572,
|
| 555 |
+
"n_pairs": 103
|
| 556 |
+
},
|
| 557 |
+
"multilingual": {
|
| 558 |
+
"rank": 0.31725888324873097,
|
| 559 |
+
"z": 3.4602457260722437,
|
| 560 |
+
"n_pairs": 394
|
| 561 |
+
},
|
| 562 |
+
"order-ops": {
|
| 563 |
+
"rank": 0.5844155844155845,
|
| 564 |
+
"z": 5.2574653549627826,
|
| 565 |
+
"n_pairs": 110
|
| 566 |
+
},
|
| 567 |
+
"poetry": {
|
| 568 |
+
"rank": 0.061224489795918366,
|
| 569 |
+
"z": 2.5307455500777887,
|
| 570 |
+
"n_pairs": 98
|
| 571 |
+
},
|
| 572 |
+
"typo": {
|
| 573 |
+
"rank": 0.32142857142857145,
|
| 574 |
+
"z": 3.5066481679677963,
|
| 575 |
+
"n_pairs": 96
|
| 576 |
+
}
|
| 577 |
+
},
|
| 578 |
+
"J|Wikipedia_en|s2": {
|
| 579 |
+
"multihop": {
|
| 580 |
+
"rank": 0.36893203883495146,
|
| 581 |
+
"z": 3.7126931484463146,
|
| 582 |
+
"n_pairs": 103
|
| 583 |
+
},
|
| 584 |
+
"multilingual": {
|
| 585 |
+
"rank": 0.3110949963741842,
|
| 586 |
+
"z": 3.4117022766377114,
|
| 587 |
+
"n_pairs": 394
|
| 588 |
+
},
|
| 589 |
+
"order-ops": {
|
| 590 |
+
"rank": 0.5805194805194805,
|
| 591 |
+
"z": 5.221943621201949,
|
| 592 |
+
"n_pairs": 110
|
| 593 |
+
},
|
| 594 |
+
"poetry": {
|
| 595 |
+
"rank": 0.06268221574344023,
|
| 596 |
+
"z": 2.513642885247055,
|
| 597 |
+
"n_pairs": 98
|
| 598 |
+
},
|
| 599 |
+
"typo": {
|
| 600 |
+
"rank": 0.3095238095238096,
|
| 601 |
+
"z": 3.487534542878469,
|
| 602 |
+
"n_pairs": 96
|
| 603 |
+
}
|
| 604 |
+
},
|
| 605 |
+
"J|USPTO_Backgrounds|s0": {
|
| 606 |
+
"multihop": {
|
| 607 |
+
"rank": 0.420249653259362,
|
| 608 |
+
"z": 3.943197340641207,
|
| 609 |
+
"n_pairs": 103
|
| 610 |
+
},
|
| 611 |
+
"multilingual": {
|
| 612 |
+
"rank": 0.3328498912255258,
|
| 613 |
+
"z": 3.617224749272245,
|
| 614 |
+
"n_pairs": 394
|
| 615 |
+
},
|
| 616 |
+
"order-ops": {
|
| 617 |
+
"rank": 0.6246753246753247,
|
| 618 |
+
"z": 5.513433169234883,
|
| 619 |
+
"n_pairs": 110
|
| 620 |
+
},
|
| 621 |
+
"poetry": {
|
| 622 |
+
"rank": 0.08454810495626822,
|
| 623 |
+
"z": 2.6456040983297386,
|
| 624 |
+
"n_pairs": 98
|
| 625 |
+
},
|
| 626 |
+
"typo": {
|
| 627 |
+
"rank": 0.42708333333333337,
|
| 628 |
+
"z": 3.7516461834311485,
|
| 629 |
+
"n_pairs": 96
|
| 630 |
+
}
|
| 631 |
+
},
|
| 632 |
+
"J|USPTO_Backgrounds|s1": {
|
| 633 |
+
"multihop": {
|
| 634 |
+
"rank": 0.4133148404993065,
|
| 635 |
+
"z": 3.9599806452260435,
|
| 636 |
+
"n_pairs": 103
|
| 637 |
+
},
|
| 638 |
+
"multilingual": {
|
| 639 |
+
"rank": 0.3390137781000725,
|
| 640 |
+
"z": 3.6621401048251214,
|
| 641 |
+
"n_pairs": 394
|
| 642 |
+
},
|
| 643 |
+
"order-ops": {
|
| 644 |
+
"rank": 0.6571428571428571,
|
| 645 |
+
"z": 5.619987377253445,
|
| 646 |
+
"n_pairs": 110
|
| 647 |
+
},
|
| 648 |
+
"poetry": {
|
| 649 |
+
"rank": 0.09329446064139943,
|
| 650 |
+
"z": 2.6892731098496183,
|
| 651 |
+
"n_pairs": 98
|
| 652 |
+
},
|
| 653 |
+
"typo": {
|
| 654 |
+
"rank": 0.44345238095238093,
|
| 655 |
+
"z": 3.7858387008309364,
|
| 656 |
+
"n_pairs": 96
|
| 657 |
+
}
|
| 658 |
+
},
|
| 659 |
+
"J|USPTO_Backgrounds|s2": {
|
| 660 |
+
"multihop": {
|
| 661 |
+
"rank": 0.39944521497919555,
|
| 662 |
+
"z": 3.9531577559350763,
|
| 663 |
+
"n_pairs": 103
|
| 664 |
+
},
|
| 665 |
+
"multilingual": {
|
| 666 |
+
"rank": 0.34590282813633066,
|
| 667 |
+
"z": 3.712495117471908,
|
| 668 |
+
"n_pairs": 394
|
| 669 |
+
},
|
| 670 |
+
"order-ops": {
|
| 671 |
+
"rank": 0.6376623376623376,
|
| 672 |
+
"z": 5.639850367199291,
|
| 673 |
+
"n_pairs": 110
|
| 674 |
+
},
|
| 675 |
+
"poetry": {
|
| 676 |
+
"rank": 0.10787172011661807,
|
| 677 |
+
"z": 2.747182385045655,
|
| 678 |
+
"n_pairs": 98
|
| 679 |
+
},
|
| 680 |
+
"typo": {
|
| 681 |
+
"rank": 0.44642857142857145,
|
| 682 |
+
"z": 3.854549737026294,
|
| 683 |
+
"n_pairs": 96
|
| 684 |
+
}
|
| 685 |
+
},
|
| 686 |
+
"J|Pile-CC|s0": {
|
| 687 |
+
"multihop": {
|
| 688 |
+
"rank": 0.3828016643550624,
|
| 689 |
+
"z": 3.903921564805855,
|
| 690 |
+
"n_pairs": 103
|
| 691 |
+
},
|
| 692 |
+
"multilingual": {
|
| 693 |
+
"rank": 0.3332124728063815,
|
| 694 |
+
"z": 3.6563817554621525,
|
| 695 |
+
"n_pairs": 394
|
| 696 |
+
},
|
| 697 |
+
"order-ops": {
|
| 698 |
+
"rank": 0.5922077922077922,
|
| 699 |
+
"z": 5.404603629762476,
|
| 700 |
+
"n_pairs": 110
|
| 701 |
+
},
|
| 702 |
+
"poetry": {
|
| 703 |
+
"rank": 0.06851311953352769,
|
| 704 |
+
"z": 2.5789884632947495,
|
| 705 |
+
"n_pairs": 98
|
| 706 |
+
},
|
| 707 |
+
"typo": {
|
| 708 |
+
"rank": 0.3630952380952381,
|
| 709 |
+
"z": 3.804555675635735,
|
| 710 |
+
"n_pairs": 96
|
| 711 |
+
}
|
| 712 |
+
},
|
| 713 |
+
"J|Pile-CC|s1": {
|
| 714 |
+
"multihop": {
|
| 715 |
+
"rank": 0.3883495145631068,
|
| 716 |
+
"z": 3.8897387078664836,
|
| 717 |
+
"n_pairs": 103
|
| 718 |
+
},
|
| 719 |
+
"multilingual": {
|
| 720 |
+
"rank": 0.3234227701232778,
|
| 721 |
+
"z": 3.56714168538902,
|
| 722 |
+
"n_pairs": 394
|
| 723 |
+
},
|
| 724 |
+
"order-ops": {
|
| 725 |
+
"rank": 0.5818181818181819,
|
| 726 |
+
"z": 5.463297904621471,
|
| 727 |
+
"n_pairs": 110
|
| 728 |
+
},
|
| 729 |
+
"poetry": {
|
| 730 |
+
"rank": 0.05393586005830903,
|
| 731 |
+
"z": 2.6064579340876364,
|
| 732 |
+
"n_pairs": 98
|
| 733 |
+
},
|
| 734 |
+
"typo": {
|
| 735 |
+
"rank": 0.40476190476190477,
|
| 736 |
+
"z": 3.8851715164879956,
|
| 737 |
+
"n_pairs": 96
|
| 738 |
+
}
|
| 739 |
+
},
|
| 740 |
+
"J|Pile-CC|s2": {
|
| 741 |
+
"multihop": {
|
| 742 |
+
"rank": 0.391123439667129,
|
| 743 |
+
"z": 3.9151969303205174,
|
| 744 |
+
"n_pairs": 103
|
| 745 |
+
},
|
| 746 |
+
"multilingual": {
|
| 747 |
+
"rank": 0.33575054387237124,
|
| 748 |
+
"z": 3.606160680805971,
|
| 749 |
+
"n_pairs": 394
|
| 750 |
+
},
|
| 751 |
+
"order-ops": {
|
| 752 |
+
"rank": 0.5909090909090909,
|
| 753 |
+
"z": 5.418469522216103,
|
| 754 |
+
"n_pairs": 110
|
| 755 |
+
},
|
| 756 |
+
"poetry": {
|
| 757 |
+
"rank": 0.06705539358600583,
|
| 758 |
+
"z": 2.5475342316286906,
|
| 759 |
+
"n_pairs": 98
|
| 760 |
+
},
|
| 761 |
+
"typo": {
|
| 762 |
+
"rank": 0.38244047619047616,
|
| 763 |
+
"z": 3.7702168126900992,
|
| 764 |
+
"n_pairs": 96
|
| 765 |
+
}
|
| 766 |
+
},
|
| 767 |
+
"J|StackExchange|s0": {
|
| 768 |
+
"multihop": {
|
| 769 |
+
"rank": 0.42718446601941745,
|
| 770 |
+
"z": 4.062260240027048,
|
| 771 |
+
"n_pairs": 103
|
| 772 |
+
},
|
| 773 |
+
"multilingual": {
|
| 774 |
+
"rank": 0.3353879622915156,
|
| 775 |
+
"z": 3.7145396611714725,
|
| 776 |
+
"n_pairs": 394
|
| 777 |
+
},
|
| 778 |
+
"order-ops": {
|
| 779 |
+
"rank": 0.5675324675324674,
|
| 780 |
+
"z": 5.500007843971252,
|
| 781 |
+
"n_pairs": 110
|
| 782 |
+
},
|
| 783 |
+
"poetry": {
|
| 784 |
+
"rank": 0.05830903790087464,
|
| 785 |
+
"z": 2.4862898527359474,
|
| 786 |
+
"n_pairs": 98
|
| 787 |
+
},
|
| 788 |
+
"typo": {
|
| 789 |
+
"rank": 0.40029761904761907,
|
| 790 |
+
"z": 3.856445398181677,
|
| 791 |
+
"n_pairs": 96
|
| 792 |
+
}
|
| 793 |
+
},
|
| 794 |
+
"J|StackExchange|s1": {
|
| 795 |
+
"multihop": {
|
| 796 |
+
"rank": 0.40083217753120665,
|
| 797 |
+
"z": 3.9458269378513964,
|
| 798 |
+
"n_pairs": 103
|
| 799 |
+
},
|
| 800 |
+
"multilingual": {
|
| 801 |
+
"rank": 0.34699057287889773,
|
| 802 |
+
"z": 3.717639919616244,
|
| 803 |
+
"n_pairs": 394
|
| 804 |
+
},
|
| 805 |
+
"order-ops": {
|
| 806 |
+
"rank": 0.5597402597402598,
|
| 807 |
+
"z": 5.454329214312813,
|
| 808 |
+
"n_pairs": 110
|
| 809 |
+
},
|
| 810 |
+
"poetry": {
|
| 811 |
+
"rank": 0.05830903790087464,
|
| 812 |
+
"z": 2.474340495406365,
|
| 813 |
+
"n_pairs": 98
|
| 814 |
+
},
|
| 815 |
+
"typo": {
|
| 816 |
+
"rank": 0.40625,
|
| 817 |
+
"z": 3.8580934616426625,
|
| 818 |
+
"n_pairs": 96
|
| 819 |
+
}
|
| 820 |
+
},
|
| 821 |
+
"J|StackExchange|s2": {
|
| 822 |
+
"multihop": {
|
| 823 |
+
"rank": 0.42857142857142866,
|
| 824 |
+
"z": 4.065570627601401,
|
| 825 |
+
"n_pairs": 103
|
| 826 |
+
},
|
| 827 |
+
"multilingual": {
|
| 828 |
+
"rank": 0.3328498912255258,
|
| 829 |
+
"z": 3.6968645884603415,
|
| 830 |
+
"n_pairs": 394
|
| 831 |
+
},
|
| 832 |
+
"order-ops": {
|
| 833 |
+
"rank": 0.5857142857142856,
|
| 834 |
+
"z": 5.5429936354810545,
|
| 835 |
+
"n_pairs": 110
|
| 836 |
+
},
|
| 837 |
+
"poetry": {
|
| 838 |
+
"rank": 0.05393586005830903,
|
| 839 |
+
"z": 2.4796188942023685,
|
| 840 |
+
"n_pairs": 98
|
| 841 |
+
},
|
| 842 |
+
"typo": {
|
| 843 |
+
"rank": 0.4032738095238095,
|
| 844 |
+
"z": 3.8242418008546033,
|
| 845 |
+
"n_pairs": 96
|
| 846 |
+
}
|
| 847 |
+
},
|
| 848 |
+
"J|Github|s0": {
|
| 849 |
+
"multihop": {
|
| 850 |
+
"rank": 0.32454923717059636,
|
| 851 |
+
"z": 3.357211357181512,
|
| 852 |
+
"n_pairs": 103
|
| 853 |
+
},
|
| 854 |
+
"multilingual": {
|
| 855 |
+
"rank": 0.23567802755620013,
|
| 856 |
+
"z": 2.9966800858829226,
|
| 857 |
+
"n_pairs": 394
|
| 858 |
+
},
|
| 859 |
+
"order-ops": {
|
| 860 |
+
"rank": 0.5999999999999999,
|
| 861 |
+
"z": 5.133947640115564,
|
| 862 |
+
"n_pairs": 110
|
| 863 |
+
},
|
| 864 |
+
"poetry": {
|
| 865 |
+
"rank": 0.030612244897959186,
|
| 866 |
+
"z": 2.2744970212177353,
|
| 867 |
+
"n_pairs": 98
|
| 868 |
+
},
|
| 869 |
+
"typo": {
|
| 870 |
+
"rank": 0.2202380952380952,
|
| 871 |
+
"z": 3.176635969430208,
|
| 872 |
+
"n_pairs": 96
|
| 873 |
+
}
|
| 874 |
+
},
|
| 875 |
+
"J|Github|s1": {
|
| 876 |
+
"multihop": {
|
| 877 |
+
"rank": 0.348127600554785,
|
| 878 |
+
"z": 3.522636915873555,
|
| 879 |
+
"n_pairs": 103
|
| 880 |
+
},
|
| 881 |
+
"multilingual": {
|
| 882 |
+
"rank": 0.2646845540246555,
|
| 883 |
+
"z": 3.135660663171468,
|
| 884 |
+
"n_pairs": 394
|
| 885 |
+
},
|
| 886 |
+
"order-ops": {
|
| 887 |
+
"rank": 0.5896103896103896,
|
| 888 |
+
"z": 5.3051688779484145,
|
| 889 |
+
"n_pairs": 110
|
| 890 |
+
},
|
| 891 |
+
"poetry": {
|
| 892 |
+
"rank": 0.039358600583090375,
|
| 893 |
+
"z": 2.3509351890914294,
|
| 894 |
+
"n_pairs": 98
|
| 895 |
+
},
|
| 896 |
+
"typo": {
|
| 897 |
+
"rank": 0.27232142857142855,
|
| 898 |
+
"z": 3.3360372707247734,
|
| 899 |
+
"n_pairs": 96
|
| 900 |
+
}
|
| 901 |
+
},
|
| 902 |
+
"J|Github|s2": {
|
| 903 |
+
"multihop": {
|
| 904 |
+
"rank": 0.32316227461858527,
|
| 905 |
+
"z": 3.3917247172698235,
|
| 906 |
+
"n_pairs": 103
|
| 907 |
+
},
|
| 908 |
+
"multilingual": {
|
| 909 |
+
"rank": 0.2548948513415518,
|
| 910 |
+
"z": 3.034596077681798,
|
| 911 |
+
"n_pairs": 394
|
| 912 |
+
},
|
| 913 |
+
"order-ops": {
|
| 914 |
+
"rank": 0.5935064935064934,
|
| 915 |
+
"z": 5.225318701700731,
|
| 916 |
+
"n_pairs": 110
|
| 917 |
+
},
|
| 918 |
+
"poetry": {
|
| 919 |
+
"rank": 0.03498542274052478,
|
| 920 |
+
"z": 2.3192666671714006,
|
| 921 |
+
"n_pairs": 98
|
| 922 |
+
},
|
| 923 |
+
"typo": {
|
| 924 |
+
"rank": 0.25595238095238093,
|
| 925 |
+
"z": 3.2398867520193257,
|
| 926 |
+
"n_pairs": 96
|
| 927 |
+
}
|
| 928 |
+
}
|
| 929 |
+
},
|
| 930 |
+
"fits": {
|
| 931 |
+
"J|Wikipedia_en|s0": {
|
| 932 |
+
"n_prompts": 25,
|
| 933 |
+
"fit_seconds": 717.8,
|
| 934 |
+
"prompts_sha256": "9e54a0769bd230f55ba825a256e03c0f34730227e20dfc1b75efc797f711db71",
|
| 935 |
+
"corpus": "Wikipedia_en",
|
| 936 |
+
"seed_block": 0,
|
| 937 |
+
"fro_norm_per_layer": {
|
| 938 |
+
"12": 25.47411346435547,
|
| 939 |
+
"13": 28.14723014831543,
|
| 940 |
+
"14": 30.691478729248047,
|
| 941 |
+
"15": 33.2955436706543,
|
| 942 |
+
"16": 35.893863677978516,
|
| 943 |
+
"17": 37.72172927856445,
|
| 944 |
+
"18": 39.467960357666016,
|
| 945 |
+
"19": 41.63821029663086,
|
| 946 |
+
"20": 43.01144027709961,
|
| 947 |
+
"21": 44.35385513305664,
|
| 948 |
+
"22": 45.25417709350586,
|
| 949 |
+
"23": 46.21498489379883,
|
| 950 |
+
"24": 47.24529266357422,
|
| 951 |
+
"25": 48.25273513793945,
|
| 952 |
+
"26": 49.452205657958984,
|
| 953 |
+
"27": 49.94655990600586,
|
| 954 |
+
"28": 50.0068359375,
|
| 955 |
+
"29": 50.67486572265625
|
| 956 |
+
}
|
| 957 |
+
},
|
| 958 |
+
"J|Wikipedia_en|s1": {
|
| 959 |
+
"n_prompts": 25,
|
| 960 |
+
"fit_seconds": 715.9,
|
| 961 |
+
"prompts_sha256": "2ee5bb179eac0bd1859f75821abcde75138e042cf6c0d93e091acb78931a9fad",
|
| 962 |
+
"corpus": "Wikipedia_en",
|
| 963 |
+
"seed_block": 1,
|
| 964 |
+
"fro_norm_per_layer": {
|
| 965 |
+
"12": 25.824872970581055,
|
| 966 |
+
"13": 28.51796531677246,
|
| 967 |
+
"14": 31.0869197845459,
|
| 968 |
+
"15": 33.67380142211914,
|
| 969 |
+
"16": 36.31361770629883,
|
| 970 |
+
"17": 38.17379379272461,
|
| 971 |
+
"18": 39.9384880065918,
|
| 972 |
+
"19": 42.0767707824707,
|
| 973 |
+
"20": 43.39258575439453,
|
| 974 |
+
"21": 44.7066650390625,
|
| 975 |
+
"22": 45.6107063293457,
|
| 976 |
+
"23": 46.55977249145508,
|
| 977 |
+
"24": 47.58478927612305,
|
| 978 |
+
"25": 48.54484939575195,
|
| 979 |
+
"26": 49.71868133544922,
|
| 980 |
+
"27": 50.12144088745117,
|
| 981 |
+
"28": 50.1476936340332,
|
| 982 |
+
"29": 50.75876235961914
|
| 983 |
+
}
|
| 984 |
+
},
|
| 985 |
+
"J|Wikipedia_en|s2": {
|
| 986 |
+
"n_prompts": 25,
|
| 987 |
+
"fit_seconds": 716.5,
|
| 988 |
+
"prompts_sha256": "dcd7a38d916ab5bc0adf8bff4400eb59938b7991e201e082cdc0164d7642659b",
|
| 989 |
+
"corpus": "Wikipedia_en",
|
| 990 |
+
"seed_block": 2,
|
| 991 |
+
"fro_norm_per_layer": {
|
| 992 |
+
"12": 25.764179229736328,
|
| 993 |
+
"13": 28.572370529174805,
|
| 994 |
+
"14": 31.217775344848633,
|
| 995 |
+
"15": 33.85537338256836,
|
| 996 |
+
"16": 36.57388687133789,
|
| 997 |
+
"17": 38.43079376220703,
|
| 998 |
+
"18": 40.218143463134766,
|
| 999 |
+
"19": 42.400333404541016,
|
| 1000 |
+
"20": 43.739261627197266,
|
| 1001 |
+
"21": 45.03773880004883,
|
| 1002 |
+
"22": 45.866294860839844,
|
| 1003 |
+
"23": 46.75215530395508,
|
| 1004 |
+
"24": 47.74189376831055,
|
| 1005 |
+
"25": 48.71947479248047,
|
| 1006 |
+
"26": 49.884422302246094,
|
| 1007 |
+
"27": 50.27415466308594,
|
| 1008 |
+
"28": 50.24495315551758,
|
| 1009 |
+
"29": 50.82538604736328
|
| 1010 |
+
}
|
| 1011 |
+
},
|
| 1012 |
+
"J|USPTO_Backgrounds|s0": {
|
| 1013 |
+
"n_prompts": 25,
|
| 1014 |
+
"fit_seconds": 715.8,
|
| 1015 |
+
"prompts_sha256": "5a8f7c90631d3924019d98c6c3a93febb4702b0267d73e3bc369a8a1e7b824e4",
|
| 1016 |
+
"corpus": "USPTO_Backgrounds",
|
| 1017 |
+
"seed_block": 0,
|
| 1018 |
+
"fro_norm_per_layer": {
|
| 1019 |
+
"12": 30.236732482910156,
|
| 1020 |
+
"13": 32.75661087036133,
|
| 1021 |
+
"14": 35.58098220825195,
|
| 1022 |
+
"15": 38.3680305480957,
|
| 1023 |
+
"16": 40.91820526123047,
|
| 1024 |
+
"17": 42.26591873168945,
|
| 1025 |
+
"18": 43.79016876220703,
|
| 1026 |
+
"19": 45.786720275878906,
|
| 1027 |
+
"20": 46.8199577331543,
|
| 1028 |
+
"21": 47.73601531982422,
|
| 1029 |
+
"22": 48.22069549560547,
|
| 1030 |
+
"23": 48.858943939208984,
|
| 1031 |
+
"24": 49.517452239990234,
|
| 1032 |
+
"25": 50.26908493041992,
|
| 1033 |
+
"26": 51.17799377441406,
|
| 1034 |
+
"27": 51.0509033203125,
|
| 1035 |
+
"28": 50.70587921142578,
|
| 1036 |
+
"29": 51.15618133544922
|
| 1037 |
+
}
|
| 1038 |
+
},
|
| 1039 |
+
"J|USPTO_Backgrounds|s1": {
|
| 1040 |
+
"n_prompts": 25,
|
| 1041 |
+
"fit_seconds": 714.0,
|
| 1042 |
+
"prompts_sha256": "2ec94c2fa4eb9cb721a09d8f1ecd51be4392b6c77ff95f9b66fb173812f04aa3",
|
| 1043 |
+
"corpus": "USPTO_Backgrounds",
|
| 1044 |
+
"seed_block": 1,
|
| 1045 |
+
"fro_norm_per_layer": {
|
| 1046 |
+
"12": 30.12862777709961,
|
| 1047 |
+
"13": 32.7627067565918,
|
| 1048 |
+
"14": 35.53038024902344,
|
| 1049 |
+
"15": 38.333805084228516,
|
| 1050 |
+
"16": 40.84742736816406,
|
| 1051 |
+
"17": 42.114013671875,
|
| 1052 |
+
"18": 43.58375930786133,
|
| 1053 |
+
"19": 45.52498245239258,
|
| 1054 |
+
"20": 46.5215950012207,
|
| 1055 |
+
"21": 47.474063873291016,
|
| 1056 |
+
"22": 47.912315368652344,
|
| 1057 |
+
"23": 48.55392837524414,
|
| 1058 |
+
"24": 49.21796798706055,
|
| 1059 |
+
"25": 49.9576416015625,
|
| 1060 |
+
"26": 50.92783737182617,
|
| 1061 |
+
"27": 50.82968521118164,
|
| 1062 |
+
"28": 50.54206085205078,
|
| 1063 |
+
"29": 51.04257583618164
|
| 1064 |
+
}
|
| 1065 |
+
},
|
| 1066 |
+
"J|USPTO_Backgrounds|s2": {
|
| 1067 |
+
"n_prompts": 25,
|
| 1068 |
+
"fit_seconds": 713.8,
|
| 1069 |
+
"prompts_sha256": "97174f19e4276737d5949b098c9ee4c669f51f178d9d499b1a5c86b6faf4afad",
|
| 1070 |
+
"corpus": "USPTO_Backgrounds",
|
| 1071 |
+
"seed_block": 2,
|
| 1072 |
+
"fro_norm_per_layer": {
|
| 1073 |
+
"12": 31.08220863342285,
|
| 1074 |
+
"13": 33.626853942871094,
|
| 1075 |
+
"14": 36.36945724487305,
|
| 1076 |
+
"15": 39.220069885253906,
|
| 1077 |
+
"16": 41.716712951660156,
|
| 1078 |
+
"17": 42.856788635253906,
|
| 1079 |
+
"18": 44.255191802978516,
|
| 1080 |
+
"19": 46.21339416503906,
|
| 1081 |
+
"20": 47.17402648925781,
|
| 1082 |
+
"21": 48.00918960571289,
|
| 1083 |
+
"22": 48.380096435546875,
|
| 1084 |
+
"23": 48.957420349121094,
|
| 1085 |
+
"24": 49.573543548583984,
|
| 1086 |
+
"25": 50.27798843383789,
|
| 1087 |
+
"26": 51.223289489746094,
|
| 1088 |
+
"27": 51.076263427734375,
|
| 1089 |
+
"28": 50.735076904296875,
|
| 1090 |
+
"29": 51.1742057800293
|
| 1091 |
+
}
|
| 1092 |
+
},
|
| 1093 |
+
"J|Pile-CC|s0": {
|
| 1094 |
+
"n_prompts": 25,
|
| 1095 |
+
"fit_seconds": 712.9,
|
| 1096 |
+
"prompts_sha256": "e0757ef16e1a8a552d467cf839e49c86eb7e90c7f090efa560e063141bf88c8e",
|
| 1097 |
+
"corpus": "Pile-CC",
|
| 1098 |
+
"seed_block": 0,
|
| 1099 |
+
"fro_norm_per_layer": {
|
| 1100 |
+
"12": 27.326833724975586,
|
| 1101 |
+
"13": 30.332229614257812,
|
| 1102 |
+
"14": 33.242000579833984,
|
| 1103 |
+
"15": 36.10725784301758,
|
| 1104 |
+
"16": 38.962711334228516,
|
| 1105 |
+
"17": 40.80857849121094,
|
| 1106 |
+
"18": 42.63570022583008,
|
| 1107 |
+
"19": 44.908409118652344,
|
| 1108 |
+
"20": 46.073490142822266,
|
| 1109 |
+
"21": 47.05865478515625,
|
| 1110 |
+
"22": 47.51934051513672,
|
| 1111 |
+
"23": 48.309513092041016,
|
| 1112 |
+
"24": 49.009803771972656,
|
| 1113 |
+
"25": 49.740360260009766,
|
| 1114 |
+
"26": 50.79483413696289,
|
| 1115 |
+
"27": 50.875579833984375,
|
| 1116 |
+
"28": 50.6129264831543,
|
| 1117 |
+
"29": 51.0517692565918
|
| 1118 |
+
}
|
| 1119 |
+
},
|
| 1120 |
+
"J|Pile-CC|s1": {
|
| 1121 |
+
"n_prompts": 25,
|
| 1122 |
+
"fit_seconds": 712.5,
|
| 1123 |
+
"prompts_sha256": "3dfe7843f22cef9b3d308d1f1cec3cade456f380b53ba1b8db8776ebe7fed77c",
|
| 1124 |
+
"corpus": "Pile-CC",
|
| 1125 |
+
"seed_block": 1,
|
| 1126 |
+
"fro_norm_per_layer": {
|
| 1127 |
+
"12": 28.01110076904297,
|
| 1128 |
+
"13": 30.7965030670166,
|
| 1129 |
+
"14": 33.55686569213867,
|
| 1130 |
+
"15": 36.41100311279297,
|
| 1131 |
+
"16": 39.08540344238281,
|
| 1132 |
+
"17": 40.86274337768555,
|
| 1133 |
+
"18": 42.52493667602539,
|
| 1134 |
+
"19": 44.74737548828125,
|
| 1135 |
+
"20": 45.96009063720703,
|
| 1136 |
+
"21": 46.924781799316406,
|
| 1137 |
+
"22": 47.40066146850586,
|
| 1138 |
+
"23": 48.23330307006836,
|
| 1139 |
+
"24": 48.95878982543945,
|
| 1140 |
+
"25": 49.670536041259766,
|
| 1141 |
+
"26": 50.714534759521484,
|
| 1142 |
+
"27": 50.77128601074219,
|
| 1143 |
+
"28": 50.543941497802734,
|
| 1144 |
+
"29": 50.97682189941406
|
| 1145 |
+
}
|
| 1146 |
+
},
|
| 1147 |
+
"J|Pile-CC|s2": {
|
| 1148 |
+
"n_prompts": 25,
|
| 1149 |
+
"fit_seconds": 712.7,
|
| 1150 |
+
"prompts_sha256": "782515fa0a56b1398a7237cda29cfac2a4e75c3bf91d1ffba99c86a19d3ab5aa",
|
| 1151 |
+
"corpus": "Pile-CC",
|
| 1152 |
+
"seed_block": 2,
|
| 1153 |
+
"fro_norm_per_layer": {
|
| 1154 |
+
"12": 27.110673904418945,
|
| 1155 |
+
"13": 30.076528549194336,
|
| 1156 |
+
"14": 32.917598724365234,
|
| 1157 |
+
"15": 35.8157844543457,
|
| 1158 |
+
"16": 38.64012908935547,
|
| 1159 |
+
"17": 40.50718307495117,
|
| 1160 |
+
"18": 42.25164794921875,
|
| 1161 |
+
"19": 44.48286437988281,
|
| 1162 |
+
"20": 45.72807312011719,
|
| 1163 |
+
"21": 46.73719787597656,
|
| 1164 |
+
"22": 47.23435592651367,
|
| 1165 |
+
"23": 48.03388214111328,
|
| 1166 |
+
"24": 48.77348327636719,
|
| 1167 |
+
"25": 49.5684928894043,
|
| 1168 |
+
"26": 50.67505645751953,
|
| 1169 |
+
"27": 50.80683898925781,
|
| 1170 |
+
"28": 50.57590866088867,
|
| 1171 |
+
"29": 51.04689025878906
|
| 1172 |
+
}
|
| 1173 |
+
},
|
| 1174 |
+
"J|StackExchange|s0": {
|
| 1175 |
+
"n_prompts": 25,
|
| 1176 |
+
"fit_seconds": 714.2,
|
| 1177 |
+
"prompts_sha256": "fb838aed6ad4bb5655506149821e49a7be3de1e40422abf3e744afa0fd80df39",
|
| 1178 |
+
"corpus": "StackExchange",
|
| 1179 |
+
"seed_block": 0,
|
| 1180 |
+
"fro_norm_per_layer": {
|
| 1181 |
+
"12": 29.31014633178711,
|
| 1182 |
+
"13": 31.878799438476562,
|
| 1183 |
+
"14": 34.66938400268555,
|
| 1184 |
+
"15": 37.237892150878906,
|
| 1185 |
+
"16": 39.83375930786133,
|
| 1186 |
+
"17": 41.555850982666016,
|
| 1187 |
+
"18": 43.22056198120117,
|
| 1188 |
+
"19": 45.463809967041016,
|
| 1189 |
+
"20": 46.5365104675293,
|
| 1190 |
+
"21": 47.511634826660156,
|
| 1191 |
+
"22": 48.09621047973633,
|
| 1192 |
+
"23": 48.869789123535156,
|
| 1193 |
+
"24": 49.59775161743164,
|
| 1194 |
+
"25": 50.24106979370117,
|
| 1195 |
+
"26": 51.18525695800781,
|
| 1196 |
+
"27": 50.98370361328125,
|
| 1197 |
+
"28": 50.5977897644043,
|
| 1198 |
+
"29": 50.98619079589844
|
| 1199 |
+
}
|
| 1200 |
+
},
|
| 1201 |
+
"J|StackExchange|s1": {
|
| 1202 |
+
"n_prompts": 25,
|
| 1203 |
+
"fit_seconds": 714.2,
|
| 1204 |
+
"prompts_sha256": "99d6fe9822aacd471eebbf96e115750a6247e2e1bef7353640f9461dfe1aad6b",
|
| 1205 |
+
"corpus": "StackExchange",
|
| 1206 |
+
"seed_block": 1,
|
| 1207 |
+
"fro_norm_per_layer": {
|
| 1208 |
+
"12": 29.9813289642334,
|
| 1209 |
+
"13": 32.756526947021484,
|
| 1210 |
+
"14": 35.454124450683594,
|
| 1211 |
+
"15": 37.913490295410156,
|
| 1212 |
+
"16": 40.541751861572266,
|
| 1213 |
+
"17": 42.225624084472656,
|
| 1214 |
+
"18": 43.83034896850586,
|
| 1215 |
+
"19": 46.03562545776367,
|
| 1216 |
+
"20": 47.06637954711914,
|
| 1217 |
+
"21": 47.918819427490234,
|
| 1218 |
+
"22": 48.48906707763672,
|
| 1219 |
+
"23": 49.20771408081055,
|
| 1220 |
+
"24": 49.94163131713867,
|
| 1221 |
+
"25": 50.54900360107422,
|
| 1222 |
+
"26": 51.485008239746094,
|
| 1223 |
+
"27": 51.22560501098633,
|
| 1224 |
+
"28": 50.785186767578125,
|
| 1225 |
+
"29": 51.12050247192383
|
| 1226 |
+
}
|
| 1227 |
+
},
|
| 1228 |
+
"J|StackExchange|s2": {
|
| 1229 |
+
"n_prompts": 25,
|
| 1230 |
+
"fit_seconds": 712.4,
|
| 1231 |
+
"prompts_sha256": "094be840f776043c464c6dc2ada90ee60b7a6349a4faf5edbee6f7646e048e58",
|
| 1232 |
+
"corpus": "StackExchange",
|
| 1233 |
+
"seed_block": 2,
|
| 1234 |
+
"fro_norm_per_layer": {
|
| 1235 |
+
"12": 29.62559700012207,
|
| 1236 |
+
"13": 32.32237243652344,
|
| 1237 |
+
"14": 35.15523910522461,
|
| 1238 |
+
"15": 37.623905181884766,
|
| 1239 |
+
"16": 40.21584701538086,
|
| 1240 |
+
"17": 41.93696975708008,
|
| 1241 |
+
"18": 43.60802459716797,
|
| 1242 |
+
"19": 45.85477828979492,
|
| 1243 |
+
"20": 46.900814056396484,
|
| 1244 |
+
"21": 47.83250045776367,
|
| 1245 |
+
"22": 48.398345947265625,
|
| 1246 |
+
"23": 49.1341667175293,
|
| 1247 |
+
"24": 49.83674621582031,
|
| 1248 |
+
"25": 50.47686767578125,
|
| 1249 |
+
"26": 51.41249084472656,
|
| 1250 |
+
"27": 51.16120529174805,
|
| 1251 |
+
"28": 50.70438003540039,
|
| 1252 |
+
"29": 51.053829193115234
|
| 1253 |
+
}
|
| 1254 |
+
},
|
| 1255 |
+
"J|Github|s0": {
|
| 1256 |
+
"n_prompts": 25,
|
| 1257 |
+
"fit_seconds": 712.5,
|
| 1258 |
+
"prompts_sha256": "fb6856c864ad3d738b939e1ca21b65a129f645bf18dd0221c4c9e2c1dadb7fe5",
|
| 1259 |
+
"corpus": "Github",
|
| 1260 |
+
"seed_block": 0,
|
| 1261 |
+
"fro_norm_per_layer": {
|
| 1262 |
+
"12": 25.596206665039062,
|
| 1263 |
+
"13": 27.109172821044922,
|
| 1264 |
+
"14": 29.121456146240234,
|
| 1265 |
+
"15": 30.757375717163086,
|
| 1266 |
+
"16": 32.600433349609375,
|
| 1267 |
+
"17": 34.267913818359375,
|
| 1268 |
+
"18": 35.85880661010742,
|
| 1269 |
+
"19": 37.88335037231445,
|
| 1270 |
+
"20": 39.29946517944336,
|
| 1271 |
+
"21": 40.87211608886719,
|
| 1272 |
+
"22": 42.26652145385742,
|
| 1273 |
+
"23": 43.78268051147461,
|
| 1274 |
+
"24": 45.19640350341797,
|
| 1275 |
+
"25": 46.571468353271484,
|
| 1276 |
+
"26": 48.045780181884766,
|
| 1277 |
+
"27": 48.60406494140625,
|
| 1278 |
+
"28": 49.111106872558594,
|
| 1279 |
+
"29": 50.13490676879883
|
| 1280 |
+
}
|
| 1281 |
+
},
|
| 1282 |
+
"J|Github|s1": {
|
| 1283 |
+
"n_prompts": 25,
|
| 1284 |
+
"fit_seconds": 711.2,
|
| 1285 |
+
"prompts_sha256": "8acd4ae02d98b1cb29ad26750a7f98b3f9bb842226f217fbc2ab18547092f0a1",
|
| 1286 |
+
"corpus": "Github",
|
| 1287 |
+
"seed_block": 1,
|
| 1288 |
+
"fro_norm_per_layer": {
|
| 1289 |
+
"12": 26.054004669189453,
|
| 1290 |
+
"13": 27.861385345458984,
|
| 1291 |
+
"14": 30.03280258178711,
|
| 1292 |
+
"15": 31.84650421142578,
|
| 1293 |
+
"16": 33.80059814453125,
|
| 1294 |
+
"17": 35.55202102661133,
|
| 1295 |
+
"18": 37.19427490234375,
|
| 1296 |
+
"19": 39.25769805908203,
|
| 1297 |
+
"20": 40.61973190307617,
|
| 1298 |
+
"21": 42.10810470581055,
|
| 1299 |
+
"22": 43.381038665771484,
|
| 1300 |
+
"23": 44.776756286621094,
|
| 1301 |
+
"24": 46.08184051513672,
|
| 1302 |
+
"25": 47.326202392578125,
|
| 1303 |
+
"26": 48.72494125366211,
|
| 1304 |
+
"27": 49.15113067626953,
|
| 1305 |
+
"28": 49.486331939697266,
|
| 1306 |
+
"29": 50.350563049316406
|
| 1307 |
+
}
|
| 1308 |
+
},
|
| 1309 |
+
"J|Github|s2": {
|
| 1310 |
+
"n_prompts": 25,
|
| 1311 |
+
"fit_seconds": 711.6,
|
| 1312 |
+
"prompts_sha256": "fb8e687f84bbaca1821949876dabd6beb24389556d54c1fa6d9234184037fc68",
|
| 1313 |
+
"corpus": "Github",
|
| 1314 |
+
"seed_block": 2,
|
| 1315 |
+
"fro_norm_per_layer": {
|
| 1316 |
+
"12": 25.251083374023438,
|
| 1317 |
+
"13": 26.807607650756836,
|
| 1318 |
+
"14": 28.799816131591797,
|
| 1319 |
+
"15": 30.52140235900879,
|
| 1320 |
+
"16": 32.41239929199219,
|
| 1321 |
+
"17": 34.15595245361328,
|
| 1322 |
+
"18": 35.801361083984375,
|
| 1323 |
+
"19": 37.893062591552734,
|
| 1324 |
+
"20": 39.344974517822266,
|
| 1325 |
+
"21": 40.995826721191406,
|
| 1326 |
+
"22": 42.41323471069336,
|
| 1327 |
+
"23": 43.94055938720703,
|
| 1328 |
+
"24": 45.35512924194336,
|
| 1329 |
+
"25": 46.73331069946289,
|
| 1330 |
+
"26": 48.25156784057617,
|
| 1331 |
+
"27": 48.75074768066406,
|
| 1332 |
+
"28": 49.18072509765625,
|
| 1333 |
+
"29": 50.145565032958984
|
| 1334 |
+
}
|
| 1335 |
+
}
|
| 1336 |
+
},
|
| 1337 |
+
"logit_I_pooled_rank_flatmean7": 0.2699603409374518,
|
| 1338 |
+
"provenance": {
|
| 1339 |
+
"schema_version": 1,
|
| 1340 |
+
"experiment": "CV6",
|
| 1341 |
+
"utc": "2026-08-24T03:25:21+00:00",
|
| 1342 |
+
"script": {
|
| 1343 |
+
"path": "experiments/cv6_per_family_ladder.py",
|
| 1344 |
+
"sha256": "f416c3cb55e95a02315504066479c2f2ba073c0d4676bbf09adb2f481cdb4a5b",
|
| 1345 |
+
"n_lines": 750
|
| 1346 |
+
},
|
| 1347 |
+
"git": {
|
| 1348 |
+
"commit": null,
|
| 1349 |
+
"commit_short": null,
|
| 1350 |
+
"branch": null,
|
| 1351 |
+
"dirty": false,
|
| 1352 |
+
"dirty_files": [],
|
| 1353 |
+
"dirty_diff_sha256": null,
|
| 1354 |
+
"describe": null
|
| 1355 |
+
},
|
| 1356 |
+
"argv": [
|
| 1357 |
+
"experiments/cv6_per_family_ladder.py",
|
| 1358 |
+
"--model",
|
| 1359 |
+
"2.8b",
|
| 1360 |
+
"--device",
|
| 1361 |
+
"cuda",
|
| 1362 |
+
"--cache-out",
|
| 1363 |
+
"results/cv6/ht_cache_2.8b.pt",
|
| 1364 |
+
"--save-lens-dir",
|
| 1365 |
+
"results/cv6/lenses"
|
| 1366 |
+
],
|
| 1367 |
+
"cwd": ".",
|
| 1368 |
+
"inputs": [
|
| 1369 |
+
{
|
| 1370 |
+
"path": "corpora/Wikipedia_en.jsonl",
|
| 1371 |
+
"sha256": "b30f4cef11add560f7df8ed07a613e3fcc6a4dda865975bec15f2952253bda77",
|
| 1372 |
+
"bytes": 10329196,
|
| 1373 |
+
"exists": true
|
| 1374 |
+
},
|
| 1375 |
+
{
|
| 1376 |
+
"path": "corpora/USPTO_Backgrounds.jsonl",
|
| 1377 |
+
"sha256": "b09ac74bdc8b169ee5c9bea6be481262f743947a92107164f30eca2346e3ccb9",
|
| 1378 |
+
"bytes": 10221054,
|
| 1379 |
+
"exists": true
|
| 1380 |
+
},
|
| 1381 |
+
{
|
| 1382 |
+
"path": "corpora/Pile-CC.jsonl",
|
| 1383 |
+
"sha256": "abbb9f19a6eb6f4c12fb382735832ceec271fece15cfebcf0fe838765d4b1e74",
|
| 1384 |
+
"bytes": 13512530,
|
| 1385 |
+
"exists": true
|
| 1386 |
+
},
|
| 1387 |
+
{
|
| 1388 |
+
"path": "corpora/StackExchange.jsonl",
|
| 1389 |
+
"sha256": "bc25e2d5775e9719ebcbdaa3bd115d827f72ebf11f47fedb377f8b0a69a28971",
|
| 1390 |
+
"bytes": 5980156,
|
| 1391 |
+
"exists": true
|
| 1392 |
+
},
|
| 1393 |
+
{
|
| 1394 |
+
"path": "corpora/Github.jsonl",
|
| 1395 |
+
"sha256": "f2bb459d10cad6c13a8b66c86191bc43df4393ba5b722a5315ea878fad2d1a26",
|
| 1396 |
+
"bytes": 17486378,
|
| 1397 |
+
"exists": true
|
| 1398 |
+
}
|
| 1399 |
+
],
|
| 1400 |
+
"env": {
|
| 1401 |
+
"python": "3.11.9",
|
| 1402 |
+
"torch": "2.11.0+cu128",
|
| 1403 |
+
"transformers": "5.15.1",
|
| 1404 |
+
"numpy": "2.4.6",
|
| 1405 |
+
"platform": "Linux-5.15.0-157-generic-x86_64-with-glibc2.35",
|
| 1406 |
+
"machine": "x86_64",
|
| 1407 |
+
"hostname": "<redacted for double-blind review>",
|
| 1408 |
+
"tf32": {
|
| 1409 |
+
"allow_tf32_matmul": false,
|
| 1410 |
+
"allow_tf32_cudnn": false,
|
| 1411 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 1412 |
+
}
|
| 1413 |
+
},
|
| 1414 |
+
"payload_sha256": "38cc82045d0d27b889fe4cd716dc6f5b6bed8d00cace9a81a394e4ead1f6c012"
|
| 1415 |
+
}
|
| 1416 |
+
}
|
results/cv6/cv6_c0_scorer_equivalence.json
ADDED
|
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV6 control C0 \u2014 the scorer reproduces D3 on the stored 410M operators",
|
| 3 |
+
"status": "CONTROL",
|
| 4 |
+
"model": "EleutherAI/pythia-410m-deduped",
|
| 5 |
+
"band": [
|
| 6 |
+
9,
|
| 7 |
+
10,
|
| 8 |
+
11,
|
| 9 |
+
12,
|
| 10 |
+
13,
|
| 11 |
+
14,
|
| 12 |
+
15,
|
| 13 |
+
16,
|
| 14 |
+
17,
|
| 15 |
+
18,
|
| 16 |
+
19,
|
| 17 |
+
20,
|
| 18 |
+
21
|
| 19 |
+
],
|
| 20 |
+
"device": "cpu",
|
| 21 |
+
"n_items": 449,
|
| 22 |
+
"n_pairs": 801,
|
| 23 |
+
"n_pairs_per_family": {
|
| 24 |
+
"multihop": 103,
|
| 25 |
+
"multilingual": 394,
|
| 26 |
+
"order-ops": 110,
|
| 27 |
+
"poetry": 98,
|
| 28 |
+
"typo": 96
|
| 29 |
+
},
|
| 30 |
+
"compared_against": "results/d3_corpus_by_family_410m.json",
|
| 31 |
+
"per_arm_deviation": {
|
| 32 |
+
"logit_I": {
|
| 33 |
+
"max_abs_dz": 0.0,
|
| 34 |
+
"max_abs_drank": 0.0
|
| 35 |
+
},
|
| 36 |
+
"Wikipedia_en|s0": {
|
| 37 |
+
"max_abs_dz": 0.0,
|
| 38 |
+
"max_abs_drank": 0.0
|
| 39 |
+
},
|
| 40 |
+
"Wikipedia_en|s1": {
|
| 41 |
+
"max_abs_dz": 0.0,
|
| 42 |
+
"max_abs_drank": 0.0
|
| 43 |
+
},
|
| 44 |
+
"Wikipedia_en|s2": {
|
| 45 |
+
"max_abs_dz": 0.0,
|
| 46 |
+
"max_abs_drank": 0.0
|
| 47 |
+
},
|
| 48 |
+
"USPTO_Backgrounds|s0": {
|
| 49 |
+
"max_abs_dz": 0.0,
|
| 50 |
+
"max_abs_drank": 0.0
|
| 51 |
+
},
|
| 52 |
+
"USPTO_Backgrounds|s1": {
|
| 53 |
+
"max_abs_dz": 0.0,
|
| 54 |
+
"max_abs_drank": 0.0
|
| 55 |
+
},
|
| 56 |
+
"USPTO_Backgrounds|s2": {
|
| 57 |
+
"max_abs_dz": 0.0,
|
| 58 |
+
"max_abs_drank": 0.0
|
| 59 |
+
},
|
| 60 |
+
"Pile-CC|s0": {
|
| 61 |
+
"max_abs_dz": 0.0,
|
| 62 |
+
"max_abs_drank": 0.0
|
| 63 |
+
},
|
| 64 |
+
"Pile-CC|s1": {
|
| 65 |
+
"max_abs_dz": 0.0,
|
| 66 |
+
"max_abs_drank": 0.0
|
| 67 |
+
},
|
| 68 |
+
"Pile-CC|s2": {
|
| 69 |
+
"max_abs_dz": 0.0,
|
| 70 |
+
"max_abs_drank": 0.0
|
| 71 |
+
},
|
| 72 |
+
"StackExchange|s0": {
|
| 73 |
+
"max_abs_dz": 0.0,
|
| 74 |
+
"max_abs_drank": 0.0
|
| 75 |
+
},
|
| 76 |
+
"StackExchange|s1": {
|
| 77 |
+
"max_abs_dz": 0.0,
|
| 78 |
+
"max_abs_drank": 0.0
|
| 79 |
+
},
|
| 80 |
+
"StackExchange|s2": {
|
| 81 |
+
"max_abs_dz": 0.0,
|
| 82 |
+
"max_abs_drank": 0.0
|
| 83 |
+
},
|
| 84 |
+
"Github|s0": {
|
| 85 |
+
"max_abs_dz": 0.0,
|
| 86 |
+
"max_abs_drank": 0.0
|
| 87 |
+
},
|
| 88 |
+
"Github|s1": {
|
| 89 |
+
"max_abs_dz": 0.0,
|
| 90 |
+
"max_abs_drank": 0.0
|
| 91 |
+
},
|
| 92 |
+
"Github|s2": {
|
| 93 |
+
"max_abs_dz": 0.0,
|
| 94 |
+
"max_abs_drank": 0.0
|
| 95 |
+
}
|
| 96 |
+
},
|
| 97 |
+
"worst_abs_dz": 0.0,
|
| 98 |
+
"worst_abs_drank": 0.0,
|
| 99 |
+
"required": "max|dz| <= 1e-6 and max|drank| <= 1e-9 over all 16 arms x 5 families, AND the negative control separates",
|
| 100 |
+
"negative_control": {
|
| 101 |
+
"what": "same scorer, LEGACY unstripped readout, identity arm, same D3 reference",
|
| 102 |
+
"why": "proves C0 can fail; a zero that a wrong readout also produces is not a test",
|
| 103 |
+
"max_abs_dz": 1.996877357092771,
|
| 104 |
+
"max_abs_drank": 0.30389610389610383,
|
| 105 |
+
"required": "max|dz| > 1e-3",
|
| 106 |
+
"separates": true,
|
| 107 |
+
"per_family_z_legacy": {
|
| 108 |
+
"multihop": 2.4813545022195984,
|
| 109 |
+
"multilingual": 2.652753429515713,
|
| 110 |
+
"order-ops": 3.0892464117570357,
|
| 111 |
+
"poetry": 2.5611357868325952,
|
| 112 |
+
"typo": 3.4371277851363025
|
| 113 |
+
}
|
| 114 |
+
},
|
| 115 |
+
"fires": true,
|
| 116 |
+
"what_it_does_not_cover": "nothing about the FIT path \u2014 it scores stored operators. It also says nothing about GPU scoring; D2 measured a CUDA-vs-CPU divergence of 2.774e-04 at the cell level.",
|
| 117 |
+
"provenance": {
|
| 118 |
+
"schema_version": 1,
|
| 119 |
+
"experiment": "CV6-C0",
|
| 120 |
+
"utc": "2026-08-23T23:04:06+00:00",
|
| 121 |
+
"script": {
|
| 122 |
+
"path": "experiments/cv6_per_family_ladder.py",
|
| 123 |
+
"sha256": "b7e6b609a8474750c75dbff539919282e7edd585ec2fba00321ae998e829711f",
|
| 124 |
+
"n_lines": 742
|
| 125 |
+
},
|
| 126 |
+
"git": {
|
| 127 |
+
"commit": "434bc077b023a5246e826a07cc325f79e6eeabbd",
|
| 128 |
+
"commit_short": "434bc077b023",
|
| 129 |
+
"branch": "pythia",
|
| 130 |
+
"dirty": true,
|
| 131 |
+
"dirty_files": [
|
| 132 |
+
"experiments/cv6_per_family_ladder.py",
|
| 133 |
+
"figures/fA_effect_and_null.png",
|
| 134 |
+
"figures/fB_validity.png",
|
| 135 |
+
"figures/make_audit_figures.py"
|
| 136 |
+
],
|
| 137 |
+
"dirty_diff_sha256": null,
|
| 138 |
+
"describe": "434bc07"
|
| 139 |
+
},
|
| 140 |
+
"argv": [
|
| 141 |
+
"experiments/cv6_per_family_ladder.py",
|
| 142 |
+
"--c0-reproduce-d3",
|
| 143 |
+
"--device",
|
| 144 |
+
"cpu"
|
| 145 |
+
],
|
| 146 |
+
"cwd": ".",
|
| 147 |
+
"inputs": [
|
| 148 |
+
{
|
| 149 |
+
"path": "results/d3_corpus_by_family_410m.json",
|
| 150 |
+
"sha256": "b4342917486e93290856203f3ec8e32b65d55bb2d389273674210b49c140ad1e",
|
| 151 |
+
"bytes": 19603,
|
| 152 |
+
"exists": true
|
| 153 |
+
}
|
| 154 |
+
],
|
| 155 |
+
"env": {
|
| 156 |
+
"python": "3.14.2",
|
| 157 |
+
"torch": "2.13.0",
|
| 158 |
+
"transformers": "5.14.1",
|
| 159 |
+
"numpy": "2.5.1",
|
| 160 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 161 |
+
"machine": "arm64",
|
| 162 |
+
"hostname": "<redacted for double-blind review>",
|
| 163 |
+
"tf32": {
|
| 164 |
+
"allow_tf32_matmul": false,
|
| 165 |
+
"allow_tf32_cudnn": false,
|
| 166 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 167 |
+
}
|
| 168 |
+
},
|
| 169 |
+
"payload_sha256": "ebbd6596928b4d50b4bc0a0034a5708d330f3d8481cc6c8b4511d190428dab36"
|
| 170 |
+
}
|
| 171 |
+
}
|
results/cv6/ht_cache_1.4b.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a3d5f7ecfb9c8bacbea1822275e28d3faf73940426c0456183b3c52fb4d83095
|
| 3 |
+
size 47847470
|
results/cv6/ht_cache_2.8b.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c4cabaca7f8ca0d38312ce257855d1e8870c1429a2f1c56e088608056881da37
|
| 3 |
+
size 82791805
|
results/cv6/lenses/lens_INSTREAM_Github_1.4b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5c3df30f285ca6a2632ddf5ccda1c0bbeb72b288167df1f5bf63e48aadfa3f68
|
| 3 |
+
size 109057367
|
results/cv6/lenses/lens_INSTREAM_Github_1.4b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3aade78c84bf70764a74a54018f1fa39966756618d5b74522194dc89e684d834
|
| 3 |
+
size 109057367
|
results/cv6/lenses/lens_INSTREAM_Github_1.4b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3b406658637da8ba432429bb262353698d1118b7efb531cb955328a95d5dbf01
|
| 3 |
+
size 109057367
|
results/cv6/lenses/lens_INSTREAM_Github_2.8b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:84f242a1abcd7293a4472f5a98fd7b5f3d047cae90890a31aa562dfb94decf5c
|
| 3 |
+
size 235936453
|
results/cv6/lenses/lens_INSTREAM_Github_2.8b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5f2d16e6011c4f2bdc865b401983bd9a490e98e5316ff0ae507d4b2e4096fd77
|
| 3 |
+
size 235936453
|
results/cv6/lenses/lens_INSTREAM_Github_2.8b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6734844c920ecfb5f3f18a4135a88679c6b50d6fe59f676f828917fc484fbab4
|
| 3 |
+
size 235936453
|
results/cv6/lenses/lens_INSTREAM_Pile-CC_1.4b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:28fb8e54aae2bea11a122fdba06f6c30b9601a006d282d7872afd0d43b2c3e70
|
| 3 |
+
size 109057386
|
results/cv6/lenses/lens_INSTREAM_Pile-CC_1.4b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:af172234b25addf33a0a086a1ecfd995e7914e4d1fdf21d588d74830e7073863
|
| 3 |
+
size 109057386
|
results/cv6/lenses/lens_INSTREAM_Pile-CC_1.4b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c6fb07ab61cb06a78913afb87427c9ca324f27a03b60f473edb83291c5c0ac00
|
| 3 |
+
size 109057386
|
results/cv6/lenses/lens_INSTREAM_Pile-CC_2.8b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e955f5d339ac9043539b179e56a9a0c59808955179af8e3aec85afc5ed384d73
|
| 3 |
+
size 235936477
|
results/cv6/lenses/lens_INSTREAM_Pile-CC_2.8b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ca7836772ad75f173f11aebfc423b61297da58c356e054cc67b0a1ec71416b5f
|
| 3 |
+
size 235936477
|
results/cv6/lenses/lens_INSTREAM_Pile-CC_2.8b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:545bdf63efc4706e446d87f8e813e992fcc3461a3a6ff074ed90564fb3ab06f1
|
| 3 |
+
size 235936477
|
results/cv6/lenses/lens_INSTREAM_StackExchange_1.4b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:469b8052d4a296314823529130a6bb8610870b860ef2c4ecbc2970c3b480b90a
|
| 3 |
+
size 109057564
|
results/cv6/lenses/lens_INSTREAM_StackExchange_1.4b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1a9f037db52482669dc656456f0643a2c290e246866875fae8e36de261ae8a52
|
| 3 |
+
size 109057564
|
results/cv6/lenses/lens_INSTREAM_StackExchange_1.4b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:07992cfcbb3d32fbeb1871f4aa79f003c6f43c4ba24ce977ae0b874f1f08a8b3
|
| 3 |
+
size 109057564
|
results/cv6/lenses/lens_INSTREAM_StackExchange_2.8b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c38d60e2daff2203100167c2b4862a31486e3fc8e70c972af4441438f7e6418e
|
| 3 |
+
size 235936685
|
results/cv6/lenses/lens_INSTREAM_StackExchange_2.8b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:45fc266499b5b97f6d5ef4d5671558a30f561430db0cd200870213950106b05d
|
| 3 |
+
size 235936685
|
results/cv6/lenses/lens_INSTREAM_StackExchange_2.8b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3438dec229cfcacf26aef81d273c2080373e5d8d5ca3dc43eb1d1280de351c1e
|
| 3 |
+
size 235936685
|
results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_1.4b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9d068456087b18d29fd38f37345ca881c14311fb97d80a280ec3adaf12cad3c5
|
| 3 |
+
size 109057640
|
results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_1.4b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:046eeef5dfb737093c822bec3607c8b22c56d5abb39e450d5aa0f7b4f8ff4265
|
| 3 |
+
size 109057640
|
results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_1.4b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3f0cd894627cc85caffa5e712da65069534ac996256fbd075a25e938b85e59c8
|
| 3 |
+
size 109057640
|
results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_2.8b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:09013abe933c8754d1ce9dcf55df4f4fc597bde7502ffb6641341b25e24242ce
|
| 3 |
+
size 235936781
|
results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_2.8b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:58fd8c7f4fb4d6e589cc559140d205d0fcda3a03ecaa95eaab83bc8e4ba2f743
|
| 3 |
+
size 235936781
|
results/cv6/lenses/lens_INSTREAM_USPTO_Backgrounds_2.8b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c9f632a13fb0b1d0eb26b639bfe046d27e7f54980f9df4a09595d9a0287527da
|
| 3 |
+
size 235936781
|
results/cv6/lenses/lens_INSTREAM_Wikipedia_en_1.4b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7db6b6e4605bd974a732b9b3501c060be612e0ba39c21c5a780658a1b64a6250
|
| 3 |
+
size 109057545
|
results/cv6/lenses/lens_INSTREAM_Wikipedia_en_1.4b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7fa1e2a931ca8454368ae8bcd67a13bc053c89e617b0c005f970b8ebbc813f76
|
| 3 |
+
size 109057545
|
results/cv6/lenses/lens_INSTREAM_Wikipedia_en_1.4b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b1763b5e8c3ca65bc26749be894aeaeb8961230491410e7853ffe45e41e77f4e
|
| 3 |
+
size 109057545
|
results/cv6/lenses/lens_INSTREAM_Wikipedia_en_2.8b_n25_s0.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1f3e2850461bf1e416585181c361f0a5018f163e3af388735e668f2c977a5a64
|
| 3 |
+
size 235936661
|
results/cv6/lenses/lens_INSTREAM_Wikipedia_en_2.8b_n25_s1.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:29f72e060566702cc51497528e6935786b949a095f1322a4832843ab633a861e
|
| 3 |
+
size 235936661
|
results/cv6/lenses/lens_INSTREAM_Wikipedia_en_2.8b_n25_s2.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a319a3aebd47c030cc5d3861f3455e91c80812eac3c7711a5d64a970206c9d6c
|
| 3 |
+
size 235936661
|
results/cv6_per_family_ladder.json
ADDED
|
@@ -0,0 +1,1238 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV6 \u2014 does the per-family corpus effect replicate at 1.4B and 2.8B?",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 4 |
+
"status": "PRE-REGISTERED",
|
| 5 |
+
"decision_rule_verbatim": "Evaluated PER FAMILY; the verdict is the count across the five. REPLICATES if R(f,2.8B) >= 10 on >= 4 of 5 families AND ordering tau >= 0.6 on >= 3 of 5. ATTENUATES if R(f,2.8B) < 5 on >= 4 of 5 families. UNCLEAR otherwise, including a split where some families replicate and others do not \u2014 a split is an informative outcome, not a failure. Do not pool to force a verdict. Do not add models to chase a threshold. Do not drop a family.",
|
| 6 |
+
"admitted_sets": [
|
| 7 |
+
"multihop",
|
| 8 |
+
"multilingual",
|
| 9 |
+
"order-ops",
|
| 10 |
+
"poetry",
|
| 11 |
+
"typo"
|
| 12 |
+
],
|
| 13 |
+
"corpora": [
|
| 14 |
+
"Wikipedia_en",
|
| 15 |
+
"USPTO_Backgrounds",
|
| 16 |
+
"Pile-CC",
|
| 17 |
+
"StackExchange",
|
| 18 |
+
"Github"
|
| 19 |
+
],
|
| 20 |
+
"seeds": [
|
| 21 |
+
0,
|
| 22 |
+
1,
|
| 23 |
+
2
|
| 24 |
+
],
|
| 25 |
+
"N": 25,
|
| 26 |
+
"K": [
|
| 27 |
+
1,
|
| 28 |
+
2,
|
| 29 |
+
5,
|
| 30 |
+
10,
|
| 31 |
+
20,
|
| 32 |
+
50,
|
| 33 |
+
100
|
| 34 |
+
],
|
| 35 |
+
"readout_convention": "STRIPPED (corrected), no prefix, flat-mean-7",
|
| 36 |
+
"pooling_convention": "flat pool over (item, intermediate) pairs WITHIN a family; seed SD is the RMS over the 5 corpora of each corpus's 3-seed sample SD \u2014 trap 4: the convention is named because it moves 'x seed SD' ratios ~37%",
|
| 37 |
+
"reference_410m": {
|
| 38 |
+
"source": "results/d3_corpus_by_family_410m.json (EXPLORATORY)",
|
| 39 |
+
"N": 200,
|
| 40 |
+
"order_by_z": {
|
| 41 |
+
"multihop": [
|
| 42 |
+
"Pile-CC",
|
| 43 |
+
"StackExchange",
|
| 44 |
+
"USPTO_Backgrounds",
|
| 45 |
+
"Wikipedia_en",
|
| 46 |
+
"Github"
|
| 47 |
+
],
|
| 48 |
+
"multilingual": [
|
| 49 |
+
"StackExchange",
|
| 50 |
+
"Pile-CC",
|
| 51 |
+
"USPTO_Backgrounds",
|
| 52 |
+
"Wikipedia_en",
|
| 53 |
+
"Github"
|
| 54 |
+
],
|
| 55 |
+
"order-ops": [
|
| 56 |
+
"USPTO_Backgrounds",
|
| 57 |
+
"StackExchange",
|
| 58 |
+
"Pile-CC",
|
| 59 |
+
"Wikipedia_en",
|
| 60 |
+
"Github"
|
| 61 |
+
],
|
| 62 |
+
"poetry": [
|
| 63 |
+
"Pile-CC",
|
| 64 |
+
"USPTO_Backgrounds",
|
| 65 |
+
"StackExchange",
|
| 66 |
+
"Wikipedia_en",
|
| 67 |
+
"Github"
|
| 68 |
+
],
|
| 69 |
+
"typo": [
|
| 70 |
+
"StackExchange",
|
| 71 |
+
"Pile-CC",
|
| 72 |
+
"USPTO_Backgrounds",
|
| 73 |
+
"Wikipedia_en",
|
| 74 |
+
"Github"
|
| 75 |
+
]
|
| 76 |
+
},
|
| 77 |
+
"R": {
|
| 78 |
+
"multihop": 25.36100542891225,
|
| 79 |
+
"multilingual": 19.426787847831193,
|
| 80 |
+
"order-ops": 24.632511354365725,
|
| 81 |
+
"poetry": 42.686224720210525,
|
| 82 |
+
"typo": 27.014210896207846
|
| 83 |
+
},
|
| 84 |
+
"z_pooled_seed_sd": {
|
| 85 |
+
"multihop": 0.017453481001782855,
|
| 86 |
+
"multilingual": 0.019695881058294912,
|
| 87 |
+
"order-ops": 0.012061420042855183,
|
| 88 |
+
"poetry": 0.005578110253433368,
|
| 89 |
+
"typo": 0.019451275697770446
|
| 90 |
+
}
|
| 91 |
+
},
|
| 92 |
+
"by_model": {
|
| 93 |
+
"1.4b": {
|
| 94 |
+
"model": "EleutherAI/pythia-1.4b-deduped",
|
| 95 |
+
"band": [
|
| 96 |
+
9,
|
| 97 |
+
10,
|
| 98 |
+
11,
|
| 99 |
+
12,
|
| 100 |
+
13,
|
| 101 |
+
14,
|
| 102 |
+
15,
|
| 103 |
+
16,
|
| 104 |
+
17,
|
| 105 |
+
18,
|
| 106 |
+
19,
|
| 107 |
+
20,
|
| 108 |
+
21
|
| 109 |
+
],
|
| 110 |
+
"n_layers": 24,
|
| 111 |
+
"device": "cuda",
|
| 112 |
+
"dtype": "float32",
|
| 113 |
+
"gpu": "NVIDIA H100 NVL",
|
| 114 |
+
"N": 25,
|
| 115 |
+
"n_pairs_per_family": {
|
| 116 |
+
"multihop": 103,
|
| 117 |
+
"multilingual": 394,
|
| 118 |
+
"order-ops": 110,
|
| 119 |
+
"poetry": 98,
|
| 120 |
+
"typo": 96
|
| 121 |
+
},
|
| 122 |
+
"logit_I": {
|
| 123 |
+
"multihop": {
|
| 124 |
+
"rank": 0.2496532593619972,
|
| 125 |
+
"z": 3.8530033965712613,
|
| 126 |
+
"n_pairs": 103
|
| 127 |
+
},
|
| 128 |
+
"multilingual": {
|
| 129 |
+
"rank": 0.1899927483683829,
|
| 130 |
+
"z": 3.2260746860564664,
|
| 131 |
+
"n_pairs": 394
|
| 132 |
+
},
|
| 133 |
+
"order-ops": {
|
| 134 |
+
"rank": 0.5597402597402598,
|
| 135 |
+
"z": 5.515925762870095,
|
| 136 |
+
"n_pairs": 110
|
| 137 |
+
},
|
| 138 |
+
"poetry": {
|
| 139 |
+
"rank": 0.008746355685131196,
|
| 140 |
+
"z": 2.4992136784962247,
|
| 141 |
+
"n_pairs": 98
|
| 142 |
+
},
|
| 143 |
+
"typo": {
|
| 144 |
+
"rank": 0.1889880952380952,
|
| 145 |
+
"z": 3.7020364006360373,
|
| 146 |
+
"n_pairs": 96
|
| 147 |
+
}
|
| 148 |
+
},
|
| 149 |
+
"logit_I_pooled_rank_flatmean7": 0.23942414367877324,
|
| 150 |
+
"by_family": {
|
| 151 |
+
"multihop": {
|
| 152 |
+
"per_corpus": {
|
| 153 |
+
"Wikipedia_en": {
|
| 154 |
+
"z_mean": 3.849424108332415,
|
| 155 |
+
"z_sd": 0.00971452307361606,
|
| 156 |
+
"rank_mean": 0.30883032824780393,
|
| 157 |
+
"rank_sd": 0.002118620293553323,
|
| 158 |
+
"n_seeds": 3
|
| 159 |
+
},
|
| 160 |
+
"USPTO_Backgrounds": {
|
| 161 |
+
"z_mean": 3.9101949866535595,
|
| 162 |
+
"z_sd": 0.03168250526694356,
|
| 163 |
+
"rank_mean": 0.32223763291724455,
|
| 164 |
+
"rank_sd": 0.003490445878534788,
|
| 165 |
+
"n_seeds": 3
|
| 166 |
+
},
|
| 167 |
+
"Pile-CC": {
|
| 168 |
+
"z_mean": 3.988609111617684,
|
| 169 |
+
"z_sd": 0.038253457386003685,
|
| 170 |
+
"rank_mean": 0.3226999537679149,
|
| 171 |
+
"rank_sd": 0.0028871927870542744,
|
| 172 |
+
"n_seeds": 3
|
| 173 |
+
},
|
| 174 |
+
"StackExchange": {
|
| 175 |
+
"z_mean": 4.083285220618387,
|
| 176 |
+
"z_sd": 0.019998217439994845,
|
| 177 |
+
"rank_mean": 0.39667128987517336,
|
| 178 |
+
"rank_sd": 0.012091259205383278,
|
| 179 |
+
"n_seeds": 3
|
| 180 |
+
},
|
| 181 |
+
"Github": {
|
| 182 |
+
"z_mean": 3.752807802366979,
|
| 183 |
+
"z_sd": 0.0855118574336993,
|
| 184 |
+
"rank_mean": 0.324086916319926,
|
| 185 |
+
"rank_sd": 0.018469703518666825,
|
| 186 |
+
"n_seeds": 3
|
| 187 |
+
}
|
| 188 |
+
},
|
| 189 |
+
"z_spread": 0.33047741825140786,
|
| 190 |
+
"z_pooled_seed_sd": 0.045329210368280526,
|
| 191 |
+
"R": 7.2906061139477085,
|
| 192 |
+
"rank_spread": 0.08784096162736943,
|
| 193 |
+
"rank_pooled_seed_sd": 0.010122609780171063,
|
| 194 |
+
"rank_spread_over_sd": 8.677699085016492,
|
| 195 |
+
"order_by_z": [
|
| 196 |
+
"StackExchange",
|
| 197 |
+
"Pile-CC",
|
| 198 |
+
"USPTO_Backgrounds",
|
| 199 |
+
"Wikipedia_en",
|
| 200 |
+
"Github"
|
| 201 |
+
],
|
| 202 |
+
"order_by_rank": [
|
| 203 |
+
"StackExchange",
|
| 204 |
+
"Github",
|
| 205 |
+
"Pile-CC",
|
| 206 |
+
"USPTO_Backgrounds",
|
| 207 |
+
"Wikipedia_en"
|
| 208 |
+
],
|
| 209 |
+
"n_pairs": 103,
|
| 210 |
+
"kendall_tau_vs_410m": 0.8,
|
| 211 |
+
"fp16_roundtrip_diagnostic": {
|
| 212 |
+
"R": 7.291143850680914,
|
| 213 |
+
"max_abs_dz_mean": 7.843778357141673e-05,
|
| 214 |
+
"order_unchanged": true
|
| 215 |
+
}
|
| 216 |
+
},
|
| 217 |
+
"multilingual": {
|
| 218 |
+
"per_corpus": {
|
| 219 |
+
"Wikipedia_en": {
|
| 220 |
+
"z_mean": 3.055601908067358,
|
| 221 |
+
"z_sd": 0.024883413619364386,
|
| 222 |
+
"rank_mean": 0.18624607203287408,
|
| 223 |
+
"rank_sd": 0.0065298248292193,
|
| 224 |
+
"n_seeds": 3
|
| 225 |
+
},
|
| 226 |
+
"USPTO_Backgrounds": {
|
| 227 |
+
"z_mean": 3.1024229467015982,
|
| 228 |
+
"z_sd": 0.021925288574017928,
|
| 229 |
+
"rank_mean": 0.18044476673918297,
|
| 230 |
+
"rank_sd": 0.0032497775343845316,
|
| 231 |
+
"n_seeds": 3
|
| 232 |
+
},
|
| 233 |
+
"Pile-CC": {
|
| 234 |
+
"z_mean": 3.2126307500094526,
|
| 235 |
+
"z_sd": 0.04011617460924262,
|
| 236 |
+
"rank_mean": 0.19639835629683347,
|
| 237 |
+
"rank_sd": 0.004202402381124895,
|
| 238 |
+
"n_seeds": 3
|
| 239 |
+
},
|
| 240 |
+
"StackExchange": {
|
| 241 |
+
"z_mean": 3.2308733033145542,
|
| 242 |
+
"z_sd": 0.055215016867437655,
|
| 243 |
+
"rank_mean": 0.2274595117234711,
|
| 244 |
+
"rank_sd": 0.012943366831597983,
|
| 245 |
+
"n_seeds": 3
|
| 246 |
+
},
|
| 247 |
+
"Github": {
|
| 248 |
+
"z_mean": 2.8394160118905822,
|
| 249 |
+
"z_sd": 0.09979349484462473,
|
| 250 |
+
"rank_mean": 0.14744984288131496,
|
| 251 |
+
"rank_sd": 0.017215308294004156,
|
| 252 |
+
"n_seeds": 3
|
| 253 |
+
}
|
| 254 |
+
},
|
| 255 |
+
"z_spread": 0.391457291423972,
|
| 256 |
+
"z_pooled_seed_sd": 0.05606540772822604,
|
| 257 |
+
"R": 6.982153653845515,
|
| 258 |
+
"rank_spread": 0.08000966884215613,
|
| 259 |
+
"rank_pooled_seed_sd": 0.010341735218918366,
|
| 260 |
+
"rank_spread_over_sd": 7.736580675145566,
|
| 261 |
+
"order_by_z": [
|
| 262 |
+
"StackExchange",
|
| 263 |
+
"Pile-CC",
|
| 264 |
+
"USPTO_Backgrounds",
|
| 265 |
+
"Wikipedia_en",
|
| 266 |
+
"Github"
|
| 267 |
+
],
|
| 268 |
+
"order_by_rank": [
|
| 269 |
+
"StackExchange",
|
| 270 |
+
"Pile-CC",
|
| 271 |
+
"Wikipedia_en",
|
| 272 |
+
"USPTO_Backgrounds",
|
| 273 |
+
"Github"
|
| 274 |
+
],
|
| 275 |
+
"n_pairs": 394,
|
| 276 |
+
"kendall_tau_vs_410m": 1.0,
|
| 277 |
+
"fp16_roundtrip_diagnostic": {
|
| 278 |
+
"R": 6.981554279672338,
|
| 279 |
+
"max_abs_dz_mean": 2.806884584893865e-05,
|
| 280 |
+
"order_unchanged": true
|
| 281 |
+
}
|
| 282 |
+
},
|
| 283 |
+
"order-ops": {
|
| 284 |
+
"per_corpus": {
|
| 285 |
+
"Wikipedia_en": {
|
| 286 |
+
"z_mean": 5.321406318924644,
|
| 287 |
+
"z_sd": 0.05556506407862744,
|
| 288 |
+
"rank_mean": 0.5632034632034633,
|
| 289 |
+
"rank_sd": 0.009918995010726954,
|
| 290 |
+
"n_seeds": 3
|
| 291 |
+
},
|
| 292 |
+
"USPTO_Backgrounds": {
|
| 293 |
+
"z_mean": 5.318626935193033,
|
| 294 |
+
"z_sd": 0.0436481925352706,
|
| 295 |
+
"rank_mean": 0.5831168831168831,
|
| 296 |
+
"rank_sd": 0.011902794012872206,
|
| 297 |
+
"n_seeds": 3
|
| 298 |
+
},
|
| 299 |
+
"Pile-CC": {
|
| 300 |
+
"z_mean": 5.474697200818495,
|
| 301 |
+
"z_sd": 0.0237919350357011,
|
| 302 |
+
"rank_mean": 0.5887445887445888,
|
| 303 |
+
"rank_sd": 0.0029992221776084464,
|
| 304 |
+
"n_seeds": 3
|
| 305 |
+
},
|
| 306 |
+
"StackExchange": {
|
| 307 |
+
"z_mean": 5.220926105976105,
|
| 308 |
+
"z_sd": 0.04991356678409937,
|
| 309 |
+
"rank_mean": 0.5896103896103896,
|
| 310 |
+
"rank_sd": 0.004682534123979193,
|
| 311 |
+
"n_seeds": 3
|
| 312 |
+
},
|
| 313 |
+
"Github": {
|
| 314 |
+
"z_mean": 5.383031076373476,
|
| 315 |
+
"z_sd": 0.07990919093950227,
|
| 316 |
+
"rank_mean": 0.5705627705627706,
|
| 317 |
+
"rank_sd": 0.01364151623947074,
|
| 318 |
+
"n_seeds": 3
|
| 319 |
+
}
|
| 320 |
+
},
|
| 321 |
+
"z_spread": 0.25377109484239035,
|
| 322 |
+
"z_pooled_seed_sd": 0.05373181585637624,
|
| 323 |
+
"R": 4.722920504319339,
|
| 324 |
+
"rank_spread": 0.026406926406926323,
|
| 325 |
+
"rank_pooled_seed_sd": 0.009561123282756314,
|
| 326 |
+
"rank_spread_over_sd": 2.761906276697819,
|
| 327 |
+
"order_by_z": [
|
| 328 |
+
"Pile-CC",
|
| 329 |
+
"Github",
|
| 330 |
+
"Wikipedia_en",
|
| 331 |
+
"USPTO_Backgrounds",
|
| 332 |
+
"StackExchange"
|
| 333 |
+
],
|
| 334 |
+
"order_by_rank": [
|
| 335 |
+
"StackExchange",
|
| 336 |
+
"Pile-CC",
|
| 337 |
+
"USPTO_Backgrounds",
|
| 338 |
+
"Github",
|
| 339 |
+
"Wikipedia_en"
|
| 340 |
+
],
|
| 341 |
+
"n_pairs": 110,
|
| 342 |
+
"kendall_tau_vs_410m": -0.4,
|
| 343 |
+
"fp16_roundtrip_diagnostic": {
|
| 344 |
+
"R": 4.72187977619938,
|
| 345 |
+
"max_abs_dz_mean": 0.00011707869443000618,
|
| 346 |
+
"order_unchanged": true
|
| 347 |
+
}
|
| 348 |
+
},
|
| 349 |
+
"poetry": {
|
| 350 |
+
"per_corpus": {
|
| 351 |
+
"Wikipedia_en": {
|
| 352 |
+
"z_mean": 2.321828322142971,
|
| 353 |
+
"z_sd": 0.010642859707299999,
|
| 354 |
+
"rank_mean": 0.02235179786200194,
|
| 355 |
+
"rank_sd": 0.0036685298519294196,
|
| 356 |
+
"n_seeds": 3
|
| 357 |
+
},
|
| 358 |
+
"USPTO_Backgrounds": {
|
| 359 |
+
"z_mean": 2.3693615256929075,
|
| 360 |
+
"z_sd": 0.016925142442212295,
|
| 361 |
+
"rank_mean": 0.023323615160349854,
|
| 362 |
+
"rank_sd": 0.0025248554046193535,
|
| 363 |
+
"n_seeds": 3
|
| 364 |
+
},
|
| 365 |
+
"Pile-CC": {
|
| 366 |
+
"z_mean": 2.3751317657986464,
|
| 367 |
+
"z_sd": 0.020213283926543354,
|
| 368 |
+
"rank_mean": 0.01652089407191448,
|
| 369 |
+
"rank_sd": 0.00303449854149582,
|
| 370 |
+
"n_seeds": 3
|
| 371 |
+
},
|
| 372 |
+
"StackExchange": {
|
| 373 |
+
"z_mean": 2.305089985635005,
|
| 374 |
+
"z_sd": 0.02054782796446748,
|
| 375 |
+
"rank_mean": 0.020894071914480075,
|
| 376 |
+
"rank_sd": 0.0016832369364129039,
|
| 377 |
+
"n_seeds": 3
|
| 378 |
+
},
|
| 379 |
+
"Github": {
|
| 380 |
+
"z_mean": 2.069341406226158,
|
| 381 |
+
"z_sd": 0.029947195245575223,
|
| 382 |
+
"rank_mean": 0.013605442176870748,
|
| 383 |
+
"rank_sd": 0.002226713165673389,
|
| 384 |
+
"n_seeds": 3
|
| 385 |
+
}
|
| 386 |
+
},
|
| 387 |
+
"z_spread": 0.30579035957248824,
|
| 388 |
+
"z_pooled_seed_sd": 0.020626950786131512,
|
| 389 |
+
"R": 14.824797069767857,
|
| 390 |
+
"rank_spread": 0.009718172983479106,
|
| 391 |
+
"rank_pooled_seed_sd": 0.002714138006563447,
|
| 392 |
+
"rank_spread_over_sd": 3.580574370197166,
|
| 393 |
+
"order_by_z": [
|
| 394 |
+
"Pile-CC",
|
| 395 |
+
"USPTO_Backgrounds",
|
| 396 |
+
"Wikipedia_en",
|
| 397 |
+
"StackExchange",
|
| 398 |
+
"Github"
|
| 399 |
+
],
|
| 400 |
+
"order_by_rank": [
|
| 401 |
+
"USPTO_Backgrounds",
|
| 402 |
+
"Wikipedia_en",
|
| 403 |
+
"StackExchange",
|
| 404 |
+
"Pile-CC",
|
| 405 |
+
"Github"
|
| 406 |
+
],
|
| 407 |
+
"n_pairs": 98,
|
| 408 |
+
"kendall_tau_vs_410m": 0.8,
|
| 409 |
+
"fp16_roundtrip_diagnostic": {
|
| 410 |
+
"R": 14.82280843880369,
|
| 411 |
+
"max_abs_dz_mean": 3.022188637524792e-05,
|
| 412 |
+
"order_unchanged": true
|
| 413 |
+
}
|
| 414 |
+
},
|
| 415 |
+
"typo": {
|
| 416 |
+
"per_corpus": {
|
| 417 |
+
"Wikipedia_en": {
|
| 418 |
+
"z_mean": 3.6687574258281126,
|
| 419 |
+
"z_sd": 0.005040123266250421,
|
| 420 |
+
"rank_mean": 0.3373015873015873,
|
| 421 |
+
"rank_sd": 0.0022731030232916094,
|
| 422 |
+
"n_seeds": 3
|
| 423 |
+
},
|
| 424 |
+
"USPTO_Backgrounds": {
|
| 425 |
+
"z_mean": 3.990972049534321,
|
| 426 |
+
"z_sd": 0.049932129992485516,
|
| 427 |
+
"rank_mean": 0.44990079365079366,
|
| 428 |
+
"rank_sd": 0.008461667712912667,
|
| 429 |
+
"n_seeds": 3
|
| 430 |
+
},
|
| 431 |
+
"Pile-CC": {
|
| 432 |
+
"z_mean": 4.020524656607045,
|
| 433 |
+
"z_sd": 0.013975432149536227,
|
| 434 |
+
"rank_mean": 0.4047619047619047,
|
| 435 |
+
"rank_sd": 0.005952380952380959,
|
| 436 |
+
"n_seeds": 3
|
| 437 |
+
},
|
| 438 |
+
"StackExchange": {
|
| 439 |
+
"z_mean": 3.995124079287052,
|
| 440 |
+
"z_sd": 0.010104712554199926,
|
| 441 |
+
"rank_mean": 0.4618055555555555,
|
| 442 |
+
"rank_sd": 0.0022731030232915794,
|
| 443 |
+
"n_seeds": 3
|
| 444 |
+
},
|
| 445 |
+
"Github": {
|
| 446 |
+
"z_mean": 3.2690880886382527,
|
| 447 |
+
"z_sd": 0.08830157076926536,
|
| 448 |
+
"rank_mean": 0.23015873015873017,
|
| 449 |
+
"rank_sd": 0.025004133256207473,
|
| 450 |
+
"n_seeds": 3
|
| 451 |
+
}
|
| 452 |
+
},
|
| 453 |
+
"z_spread": 0.7514365679687924,
|
| 454 |
+
"z_pooled_seed_sd": 0.04607212990087249,
|
| 455 |
+
"R": 16.31000280615553,
|
| 456 |
+
"rank_spread": 0.23164682539682535,
|
| 457 |
+
"rank_pooled_seed_sd": 0.012186642967521382,
|
| 458 |
+
"rank_spread_over_sd": 19.0082556791224,
|
| 459 |
+
"order_by_z": [
|
| 460 |
+
"Pile-CC",
|
| 461 |
+
"StackExchange",
|
| 462 |
+
"USPTO_Backgrounds",
|
| 463 |
+
"Wikipedia_en",
|
| 464 |
+
"Github"
|
| 465 |
+
],
|
| 466 |
+
"order_by_rank": [
|
| 467 |
+
"StackExchange",
|
| 468 |
+
"USPTO_Backgrounds",
|
| 469 |
+
"Pile-CC",
|
| 470 |
+
"Wikipedia_en",
|
| 471 |
+
"Github"
|
| 472 |
+
],
|
| 473 |
+
"n_pairs": 96,
|
| 474 |
+
"kendall_tau_vs_410m": 0.8,
|
| 475 |
+
"fp16_roundtrip_diagnostic": {
|
| 476 |
+
"R": 16.30612241371698,
|
| 477 |
+
"max_abs_dz_mean": 2.035333050631749e-05,
|
| 478 |
+
"order_unchanged": true
|
| 479 |
+
}
|
| 480 |
+
}
|
| 481 |
+
}
|
| 482 |
+
},
|
| 483 |
+
"2.8b": {
|
| 484 |
+
"model": "EleutherAI/pythia-2.8b-deduped",
|
| 485 |
+
"band": [
|
| 486 |
+
12,
|
| 487 |
+
13,
|
| 488 |
+
14,
|
| 489 |
+
15,
|
| 490 |
+
16,
|
| 491 |
+
17,
|
| 492 |
+
18,
|
| 493 |
+
19,
|
| 494 |
+
20,
|
| 495 |
+
21,
|
| 496 |
+
22,
|
| 497 |
+
23,
|
| 498 |
+
24,
|
| 499 |
+
25,
|
| 500 |
+
26,
|
| 501 |
+
27,
|
| 502 |
+
28,
|
| 503 |
+
29
|
| 504 |
+
],
|
| 505 |
+
"n_layers": 32,
|
| 506 |
+
"device": "cuda",
|
| 507 |
+
"dtype": "float32",
|
| 508 |
+
"gpu": "NVIDIA H100 NVL",
|
| 509 |
+
"N": 25,
|
| 510 |
+
"n_pairs_per_family": {
|
| 511 |
+
"multihop": 103,
|
| 512 |
+
"multilingual": 394,
|
| 513 |
+
"order-ops": 110,
|
| 514 |
+
"poetry": 98,
|
| 515 |
+
"typo": 96
|
| 516 |
+
},
|
| 517 |
+
"logit_I": {
|
| 518 |
+
"multihop": {
|
| 519 |
+
"rank": 0.27600554785020803,
|
| 520 |
+
"z": 3.693257671826094,
|
| 521 |
+
"n_pairs": 103
|
| 522 |
+
},
|
| 523 |
+
"multilingual": {
|
| 524 |
+
"rank": 0.22588832487309643,
|
| 525 |
+
"z": 3.0744364978603755,
|
| 526 |
+
"n_pairs": 394
|
| 527 |
+
},
|
| 528 |
+
"order-ops": {
|
| 529 |
+
"rank": 0.538961038961039,
|
| 530 |
+
"z": 5.5365345900708975,
|
| 531 |
+
"n_pairs": 110
|
| 532 |
+
},
|
| 533 |
+
"poetry": {
|
| 534 |
+
"rank": 0.02769679300291545,
|
| 535 |
+
"z": 2.430429516398177,
|
| 536 |
+
"n_pairs": 98
|
| 537 |
+
},
|
| 538 |
+
"typo": {
|
| 539 |
+
"rank": 0.28125,
|
| 540 |
+
"z": 3.6758319437503815,
|
| 541 |
+
"n_pairs": 96
|
| 542 |
+
}
|
| 543 |
+
},
|
| 544 |
+
"logit_I_pooled_rank_flatmean7": 0.2699603409374518,
|
| 545 |
+
"by_family": {
|
| 546 |
+
"multihop": {
|
| 547 |
+
"per_corpus": {
|
| 548 |
+
"Wikipedia_en": {
|
| 549 |
+
"z_mean": 3.7481867645936493,
|
| 550 |
+
"z_sd": 0.03554648642712381,
|
| 551 |
+
"rank_mean": 0.37216828478964403,
|
| 552 |
+
"rank_sd": 0.003490445878534788,
|
| 553 |
+
"n_seeds": 3
|
| 554 |
+
},
|
| 555 |
+
"USPTO_Backgrounds": {
|
| 556 |
+
"z_mean": 3.9520995335285716,
|
| 557 |
+
"z_sd": 0.008418421244240502,
|
| 558 |
+
"rank_mean": 0.4110032362459547,
|
| 559 |
+
"rank_sd": 0.010593101467766616,
|
| 560 |
+
"n_seeds": 3
|
| 561 |
+
},
|
| 562 |
+
"Pile-CC": {
|
| 563 |
+
"z_mean": 3.90297069981646,
|
| 564 |
+
"z_sd": 0.01286317636152229,
|
| 565 |
+
"rank_mean": 0.3874248728617661,
|
| 566 |
+
"rank_sd": 0.004237240587106646,
|
| 567 |
+
"n_seeds": 3
|
| 568 |
+
},
|
| 569 |
+
"StackExchange": {
|
| 570 |
+
"z_mean": 4.024555764537799,
|
| 571 |
+
"z_sd": 0.06825011963705907,
|
| 572 |
+
"rank_mean": 0.41886269070735094,
|
| 573 |
+
"rank_sd": 0.0156302741603116,
|
| 574 |
+
"n_seeds": 3
|
| 575 |
+
},
|
| 576 |
+
"Github": {
|
| 577 |
+
"z_mean": 3.4238723230979202,
|
| 578 |
+
"z_sd": 0.08735072241195618,
|
| 579 |
+
"rank_mean": 0.3319463707813222,
|
| 580 |
+
"rank_sd": 0.014030504766984291,
|
| 581 |
+
"n_seeds": 3
|
| 582 |
+
}
|
| 583 |
+
},
|
| 584 |
+
"z_spread": 0.6006834414398785,
|
| 585 |
+
"z_pooled_seed_sd": 0.052513067623269356,
|
| 586 |
+
"R": 11.43874217650363,
|
| 587 |
+
"rank_spread": 0.08691631992602872,
|
| 588 |
+
"rank_pooled_seed_sd": 0.010802886218722078,
|
| 589 |
+
"rank_spread_over_sd": 8.045657259205164,
|
| 590 |
+
"order_by_z": [
|
| 591 |
+
"StackExchange",
|
| 592 |
+
"USPTO_Backgrounds",
|
| 593 |
+
"Pile-CC",
|
| 594 |
+
"Wikipedia_en",
|
| 595 |
+
"Github"
|
| 596 |
+
],
|
| 597 |
+
"order_by_rank": [
|
| 598 |
+
"StackExchange",
|
| 599 |
+
"USPTO_Backgrounds",
|
| 600 |
+
"Pile-CC",
|
| 601 |
+
"Wikipedia_en",
|
| 602 |
+
"Github"
|
| 603 |
+
],
|
| 604 |
+
"n_pairs": 103,
|
| 605 |
+
"kendall_tau_vs_410m": 0.6,
|
| 606 |
+
"fp16_roundtrip_diagnostic": {
|
| 607 |
+
"R": 11.450690222634297,
|
| 608 |
+
"max_abs_dz_mean": 7.72414469798477e-05,
|
| 609 |
+
"order_unchanged": true
|
| 610 |
+
}
|
| 611 |
+
},
|
| 612 |
+
"multilingual": {
|
| 613 |
+
"per_corpus": {
|
| 614 |
+
"Wikipedia_en": {
|
| 615 |
+
"z_mean": 3.434918012060288,
|
| 616 |
+
"z_sd": 0.02439008151297054,
|
| 617 |
+
"rank_mean": 0.31532511481750064,
|
| 618 |
+
"rank_sd": 0.003667873073602506,
|
| 619 |
+
"n_seeds": 3
|
| 620 |
+
},
|
| 621 |
+
"USPTO_Backgrounds": {
|
| 622 |
+
"z_mean": 3.663950873964329,
|
| 623 |
+
"z_sd": 0.04769042760272363,
|
| 624 |
+
"rank_mean": 0.3392554991539763,
|
| 625 |
+
"rank_sd": 0.00652982482921927,
|
| 626 |
+
"n_seeds": 3
|
| 627 |
+
},
|
| 628 |
+
"Pile-CC": {
|
| 629 |
+
"z_mean": 3.6098976766094943,
|
| 630 |
+
"z_sd": 0.04474493232770196,
|
| 631 |
+
"rank_mean": 0.3307952622673435,
|
| 632 |
+
"rank_sd": 0.006509660628230673,
|
| 633 |
+
"n_seeds": 3
|
| 634 |
+
},
|
| 635 |
+
"StackExchange": {
|
| 636 |
+
"z_mean": 3.7096787147513903,
|
| 637 |
+
"z_sd": 0.011183248531158265,
|
| 638 |
+
"rank_mean": 0.3385303359922649,
|
| 639 |
+
"rank_sd": 0.007468945881305459,
|
| 640 |
+
"n_seeds": 3
|
| 641 |
+
},
|
| 642 |
+
"Github": {
|
| 643 |
+
"z_mean": 3.0556710560612266,
|
| 644 |
+
"z_sd": 0.07188945212789528,
|
| 645 |
+
"rank_mean": 0.2517524776408025,
|
| 646 |
+
"rank_sd": 0.014756372480988393,
|
| 647 |
+
"n_seeds": 3
|
| 648 |
+
}
|
| 649 |
+
},
|
| 650 |
+
"z_spread": 0.6540076586901638,
|
| 651 |
+
"z_pooled_seed_sd": 0.04508773736862208,
|
| 652 |
+
"R": 14.505222414316746,
|
| 653 |
+
"rank_spread": 0.0875030215131738,
|
| 654 |
+
"rank_pooled_seed_sd": 0.00862558134915699,
|
| 655 |
+
"rank_spread_over_sd": 10.144594082545613,
|
| 656 |
+
"order_by_z": [
|
| 657 |
+
"StackExchange",
|
| 658 |
+
"USPTO_Backgrounds",
|
| 659 |
+
"Pile-CC",
|
| 660 |
+
"Wikipedia_en",
|
| 661 |
+
"Github"
|
| 662 |
+
],
|
| 663 |
+
"order_by_rank": [
|
| 664 |
+
"USPTO_Backgrounds",
|
| 665 |
+
"StackExchange",
|
| 666 |
+
"Pile-CC",
|
| 667 |
+
"Wikipedia_en",
|
| 668 |
+
"Github"
|
| 669 |
+
],
|
| 670 |
+
"n_pairs": 394,
|
| 671 |
+
"kendall_tau_vs_410m": 0.8,
|
| 672 |
+
"fp16_roundtrip_diagnostic": {
|
| 673 |
+
"R": 14.514366209896588,
|
| 674 |
+
"max_abs_dz_mean": 5.073651242781807e-05,
|
| 675 |
+
"order_unchanged": true
|
| 676 |
+
}
|
| 677 |
+
},
|
| 678 |
+
"order-ops": {
|
| 679 |
+
"per_corpus": {
|
| 680 |
+
"Wikipedia_en": {
|
| 681 |
+
"z_mean": 5.21647632447156,
|
| 682 |
+
"z_sd": 0.04405024461019303,
|
| 683 |
+
"rank_mean": 0.577922077922078,
|
| 684 |
+
"rank_sd": 0.008110387010907067,
|
| 685 |
+
"n_seeds": 3
|
| 686 |
+
},
|
| 687 |
+
"USPTO_Backgrounds": {
|
| 688 |
+
"z_mean": 5.591100374135104,
|
| 689 |
+
"z_sd": 0.06802171764709751,
|
| 690 |
+
"rank_mean": 0.6398268398268399,
|
| 691 |
+
"rank_sd": 0.016341632976776498,
|
| 692 |
+
"n_seeds": 3
|
| 693 |
+
},
|
| 694 |
+
"Pile-CC": {
|
| 695 |
+
"z_mean": 5.428856505950292,
|
| 696 |
+
"z_sd": 0.03060876969522738,
|
| 697 |
+
"rank_mean": 0.5883116883116883,
|
| 698 |
+
"rank_sd": 0.005660907718883943,
|
| 699 |
+
"n_seeds": 3
|
| 700 |
+
},
|
| 701 |
+
"StackExchange": {
|
| 702 |
+
"z_mean": 5.4992225209871926,
|
| 703 |
+
"z_sd": 0.04431453680873469,
|
| 704 |
+
"rank_mean": 0.570995670995671,
|
| 705 |
+
"rank_sd": 0.013328834907654201,
|
| 706 |
+
"n_seeds": 3
|
| 707 |
+
},
|
| 708 |
+
"Github": {
|
| 709 |
+
"z_mean": 5.221422637592663,
|
| 710 |
+
"z_sd": 0.08568503459405673,
|
| 711 |
+
"rank_mean": 0.5943722943722943,
|
| 712 |
+
"rank_sd": 0.005248638810814727,
|
| 713 |
+
"n_seeds": 3
|
| 714 |
+
}
|
| 715 |
+
},
|
| 716 |
+
"z_spread": 0.37462404966354423,
|
| 717 |
+
"z_pooled_seed_sd": 0.057982718511798266,
|
| 718 |
+
"R": 6.460960425429451,
|
| 719 |
+
"rank_spread": 0.06883116883116891,
|
| 720 |
+
"rank_pooled_seed_sd": 0.010677820670698406,
|
| 721 |
+
"rank_spread_over_sd": 6.446181384189407,
|
| 722 |
+
"order_by_z": [
|
| 723 |
+
"USPTO_Backgrounds",
|
| 724 |
+
"StackExchange",
|
| 725 |
+
"Pile-CC",
|
| 726 |
+
"Github",
|
| 727 |
+
"Wikipedia_en"
|
| 728 |
+
],
|
| 729 |
+
"order_by_rank": [
|
| 730 |
+
"USPTO_Backgrounds",
|
| 731 |
+
"Github",
|
| 732 |
+
"Pile-CC",
|
| 733 |
+
"Wikipedia_en",
|
| 734 |
+
"StackExchange"
|
| 735 |
+
],
|
| 736 |
+
"n_pairs": 110,
|
| 737 |
+
"kendall_tau_vs_410m": 0.8,
|
| 738 |
+
"fp16_roundtrip_diagnostic": {
|
| 739 |
+
"R": 6.4624203254229515,
|
| 740 |
+
"max_abs_dz_mean": 0.00011228973215260396,
|
| 741 |
+
"order_unchanged": true
|
| 742 |
+
}
|
| 743 |
+
},
|
| 744 |
+
"poetry": {
|
| 745 |
+
"per_corpus": {
|
| 746 |
+
"Wikipedia_en": {
|
| 747 |
+
"z_mean": 2.5215682663074155,
|
| 748 |
+
"z_sd": 0.008653052690162434,
|
| 749 |
+
"rank_mean": 0.06268221574344023,
|
| 750 |
+
"rank_sd": 8.012344526598183e-18,
|
| 751 |
+
"n_seeds": 3
|
| 752 |
+
},
|
| 753 |
+
"USPTO_Backgrounds": {
|
| 754 |
+
"z_mean": 2.6940602347964333,
|
| 755 |
+
"z_sd": 0.051012720432079404,
|
| 756 |
+
"rank_mean": 0.09523809523809523,
|
| 757 |
+
"rank_sd": 0.011782658554890317,
|
| 758 |
+
"n_seeds": 3
|
| 759 |
+
},
|
| 760 |
+
"Pile-CC": {
|
| 761 |
+
"z_mean": 2.577650401462503,
|
| 762 |
+
"z_sd": 0.029429444912468822,
|
| 763 |
+
"rank_mean": 0.06316812439261418,
|
| 764 |
+
"rank_sd": 0.008028528494586155,
|
| 765 |
+
"n_seeds": 3
|
| 766 |
+
},
|
| 767 |
+
"StackExchange": {
|
| 768 |
+
"z_mean": 2.4801038943585896,
|
| 769 |
+
"z_sd": 0.005988009335835477,
|
| 770 |
+
"rank_mean": 0.05685131195335277,
|
| 771 |
+
"rank_sd": 0.0025248554046193596,
|
| 772 |
+
"n_seeds": 3
|
| 773 |
+
},
|
| 774 |
+
"Github": {
|
| 775 |
+
"z_mean": 2.3148899463569226,
|
| 776 |
+
"z_sd": 0.03841326630104487,
|
| 777 |
+
"rank_mean": 0.03498542274052478,
|
| 778 |
+
"rank_sd": 0.004373177842565594,
|
| 779 |
+
"n_seeds": 3
|
| 780 |
+
}
|
| 781 |
+
},
|
| 782 |
+
"z_spread": 0.3791702884395107,
|
| 783 |
+
"z_pooled_seed_sd": 0.031795284173340224,
|
| 784 |
+
"R": 11.925362464834903,
|
| 785 |
+
"rank_spread": 0.06025267249757045,
|
| 786 |
+
"rank_pooled_seed_sd": 0.006764434812553877,
|
| 787 |
+
"rank_spread_over_sd": 8.907273729026057,
|
| 788 |
+
"order_by_z": [
|
| 789 |
+
"USPTO_Backgrounds",
|
| 790 |
+
"Pile-CC",
|
| 791 |
+
"Wikipedia_en",
|
| 792 |
+
"StackExchange",
|
| 793 |
+
"Github"
|
| 794 |
+
],
|
| 795 |
+
"order_by_rank": [
|
| 796 |
+
"USPTO_Backgrounds",
|
| 797 |
+
"Pile-CC",
|
| 798 |
+
"Wikipedia_en",
|
| 799 |
+
"StackExchange",
|
| 800 |
+
"Github"
|
| 801 |
+
],
|
| 802 |
+
"n_pairs": 98,
|
| 803 |
+
"kendall_tau_vs_410m": 0.6,
|
| 804 |
+
"fp16_roundtrip_diagnostic": {
|
| 805 |
+
"R": 11.928150397522604,
|
| 806 |
+
"max_abs_dz_mean": 4.037038809601157e-05,
|
| 807 |
+
"order_unchanged": true
|
| 808 |
+
}
|
| 809 |
+
},
|
| 810 |
+
"typo": {
|
| 811 |
+
"per_corpus": {
|
| 812 |
+
"Wikipedia_en": {
|
| 813 |
+
"z_mean": 3.484520903064145,
|
| 814 |
+
"z_sd": 0.023824517984755707,
|
| 815 |
+
"rank_mean": 0.3090277777777778,
|
| 816 |
+
"rank_sd": 0.011168978421823135,
|
| 817 |
+
"n_seeds": 3
|
| 818 |
+
},
|
| 819 |
+
"USPTO_Backgrounds": {
|
| 820 |
+
"z_mean": 3.7973591391411095,
|
| 821 |
+
"z_sd": 0.05245448951050608,
|
| 822 |
+
"rank_mean": 0.43898809523809523,
|
| 823 |
+
"rank_sd": 0.010416666666666649,
|
| 824 |
+
"n_seeds": 3
|
| 825 |
+
},
|
| 826 |
+
"Pile-CC": {
|
| 827 |
+
"z_mean": 3.8199612237513065,
|
| 828 |
+
"z_sd": 0.058963986684690994,
|
| 829 |
+
"rank_mean": 0.38343253968253965,
|
| 830 |
+
"rank_sd": 0.02085104122716638,
|
| 831 |
+
"n_seeds": 3
|
| 832 |
+
},
|
| 833 |
+
"StackExchange": {
|
| 834 |
+
"z_mean": 3.846287661128574,
|
| 835 |
+
"z_sd": 0.019121880506788938,
|
| 836 |
+
"rank_mean": 0.40327380952380953,
|
| 837 |
+
"rank_sd": 0.0029761904761904656,
|
| 838 |
+
"n_seeds": 3
|
| 839 |
+
},
|
| 840 |
+
"Github": {
|
| 841 |
+
"z_mean": 3.250814785146051,
|
| 842 |
+
"z_sd": 0.08028488172623756,
|
| 843 |
+
"rank_mean": 0.24950396825396823,
|
| 844 |
+
"rank_sd": 0.02663371777511667,
|
| 845 |
+
"n_seeds": 3
|
| 846 |
+
}
|
| 847 |
+
},
|
| 848 |
+
"z_spread": 0.595472875982523,
|
| 849 |
+
"z_pooled_seed_sd": 0.052167310455261205,
|
| 850 |
+
"R": 11.414674645594424,
|
| 851 |
+
"rank_spread": 0.189484126984127,
|
| 852 |
+
"rank_pooled_seed_sd": 0.016650715155594485,
|
| 853 |
+
"rank_spread_over_sd": 11.379939252667,
|
| 854 |
+
"order_by_z": [
|
| 855 |
+
"StackExchange",
|
| 856 |
+
"Pile-CC",
|
| 857 |
+
"USPTO_Backgrounds",
|
| 858 |
+
"Wikipedia_en",
|
| 859 |
+
"Github"
|
| 860 |
+
],
|
| 861 |
+
"order_by_rank": [
|
| 862 |
+
"USPTO_Backgrounds",
|
| 863 |
+
"StackExchange",
|
| 864 |
+
"Pile-CC",
|
| 865 |
+
"Wikipedia_en",
|
| 866 |
+
"Github"
|
| 867 |
+
],
|
| 868 |
+
"n_pairs": 96,
|
| 869 |
+
"kendall_tau_vs_410m": 1.0,
|
| 870 |
+
"fp16_roundtrip_diagnostic": {
|
| 871 |
+
"R": 11.415194254884774,
|
| 872 |
+
"max_abs_dz_mean": 3.8545578718185425e-05,
|
| 873 |
+
"order_unchanged": true
|
| 874 |
+
}
|
| 875 |
+
}
|
| 876 |
+
}
|
| 877 |
+
}
|
| 878 |
+
},
|
| 879 |
+
"controls": {
|
| 880 |
+
"C0_scorer_equivalence": {
|
| 881 |
+
"required": "the scorer reproduces D3 on the stored 410M operators",
|
| 882 |
+
"result": {
|
| 883 |
+
"worst_abs_dz": 0.0,
|
| 884 |
+
"worst_abs_drank": 0.0,
|
| 885 |
+
"fires": true
|
| 886 |
+
},
|
| 887 |
+
"fires": true
|
| 888 |
+
},
|
| 889 |
+
"C1_logit_identity_arm": {
|
| 890 |
+
"required": "logit_I emitted for every model and family; at 410M the pooled admitted-5 flat-mean-7 min-rank equals 0.19810852520167826",
|
| 891 |
+
"emitted_per_model": {
|
| 892 |
+
"1.4b": [
|
| 893 |
+
"multihop",
|
| 894 |
+
"multilingual",
|
| 895 |
+
"order-ops",
|
| 896 |
+
"poetry",
|
| 897 |
+
"typo"
|
| 898 |
+
],
|
| 899 |
+
"2.8b": [
|
| 900 |
+
"multihop",
|
| 901 |
+
"multilingual",
|
| 902 |
+
"order-ops",
|
| 903 |
+
"poetry",
|
| 904 |
+
"typo"
|
| 905 |
+
]
|
| 906 |
+
},
|
| 907 |
+
"410m_anchor": "C0 scored the identity arm at 410M; see cv6_c0_scorer_equivalence.json",
|
| 908 |
+
"fires": true
|
| 909 |
+
},
|
| 910 |
+
"C2_band_matches_rule": {
|
| 911 |
+
"required": "emitted band == band_for(n_layers); a tuned band voids the run",
|
| 912 |
+
"per_model": {
|
| 913 |
+
"1.4b": {
|
| 914 |
+
"band": [
|
| 915 |
+
9,
|
| 916 |
+
10,
|
| 917 |
+
11,
|
| 918 |
+
12,
|
| 919 |
+
13,
|
| 920 |
+
14,
|
| 921 |
+
15,
|
| 922 |
+
16,
|
| 923 |
+
17,
|
| 924 |
+
18,
|
| 925 |
+
19,
|
| 926 |
+
20,
|
| 927 |
+
21
|
| 928 |
+
],
|
| 929 |
+
"rule": [
|
| 930 |
+
9,
|
| 931 |
+
10,
|
| 932 |
+
11,
|
| 933 |
+
12,
|
| 934 |
+
13,
|
| 935 |
+
14,
|
| 936 |
+
15,
|
| 937 |
+
16,
|
| 938 |
+
17,
|
| 939 |
+
18,
|
| 940 |
+
19,
|
| 941 |
+
20,
|
| 942 |
+
21
|
| 943 |
+
],
|
| 944 |
+
"match": true
|
| 945 |
+
},
|
| 946 |
+
"2.8b": {
|
| 947 |
+
"band": [
|
| 948 |
+
12,
|
| 949 |
+
13,
|
| 950 |
+
14,
|
| 951 |
+
15,
|
| 952 |
+
16,
|
| 953 |
+
17,
|
| 954 |
+
18,
|
| 955 |
+
19,
|
| 956 |
+
20,
|
| 957 |
+
21,
|
| 958 |
+
22,
|
| 959 |
+
23,
|
| 960 |
+
24,
|
| 961 |
+
25,
|
| 962 |
+
26,
|
| 963 |
+
27,
|
| 964 |
+
28,
|
| 965 |
+
29
|
| 966 |
+
],
|
| 967 |
+
"rule": [
|
| 968 |
+
12,
|
| 969 |
+
13,
|
| 970 |
+
14,
|
| 971 |
+
15,
|
| 972 |
+
16,
|
| 973 |
+
17,
|
| 974 |
+
18,
|
| 975 |
+
19,
|
| 976 |
+
20,
|
| 977 |
+
21,
|
| 978 |
+
22,
|
| 979 |
+
23,
|
| 980 |
+
24,
|
| 981 |
+
25,
|
| 982 |
+
26,
|
| 983 |
+
27,
|
| 984 |
+
28,
|
| 985 |
+
29
|
| 986 |
+
],
|
| 987 |
+
"match": true
|
| 988 |
+
}
|
| 989 |
+
},
|
| 990 |
+
"fires": true
|
| 991 |
+
},
|
| 992 |
+
"C3_seed_sd_nondegenerate": {
|
| 993 |
+
"required": "pooled_seed_sd_z(f, m) > 0 for every (family, model); else that cell is VOID",
|
| 994 |
+
"per_model": {
|
| 995 |
+
"1.4b": {
|
| 996 |
+
"multihop": 0.045329210368280526,
|
| 997 |
+
"multilingual": 0.05606540772822604,
|
| 998 |
+
"order-ops": 0.05373181585637624,
|
| 999 |
+
"poetry": 0.020626950786131512,
|
| 1000 |
+
"typo": 0.04607212990087249
|
| 1001 |
+
},
|
| 1002 |
+
"2.8b": {
|
| 1003 |
+
"multihop": 0.052513067623269356,
|
| 1004 |
+
"multilingual": 0.04508773736862208,
|
| 1005 |
+
"order-ops": 0.057982718511798266,
|
| 1006 |
+
"poetry": 0.031795284173340224,
|
| 1007 |
+
"typo": 0.052167310455261205
|
| 1008 |
+
}
|
| 1009 |
+
},
|
| 1010 |
+
"void_cells": [],
|
| 1011 |
+
"fires": true
|
| 1012 |
+
},
|
| 1013 |
+
"C4_N_identical": {
|
| 1014 |
+
"required": "N == 25 and identical across all 15 arms of every model",
|
| 1015 |
+
"n_prompts_seen": {
|
| 1016 |
+
"1.4b": [
|
| 1017 |
+
25
|
| 1018 |
+
],
|
| 1019 |
+
"2.8b": [
|
| 1020 |
+
25
|
| 1021 |
+
]
|
| 1022 |
+
},
|
| 1023 |
+
"n_arms": {
|
| 1024 |
+
"1.4b": 15,
|
| 1025 |
+
"2.8b": 15
|
| 1026 |
+
},
|
| 1027 |
+
"prompt_sha_identical_across_models": {
|
| 1028 |
+
"Wikipedia_en|s0": true,
|
| 1029 |
+
"Wikipedia_en|s1": true,
|
| 1030 |
+
"Wikipedia_en|s2": true,
|
| 1031 |
+
"USPTO_Backgrounds|s0": true,
|
| 1032 |
+
"USPTO_Backgrounds|s1": true,
|
| 1033 |
+
"USPTO_Backgrounds|s2": true,
|
| 1034 |
+
"Pile-CC|s0": true,
|
| 1035 |
+
"Pile-CC|s1": true,
|
| 1036 |
+
"Pile-CC|s2": true,
|
| 1037 |
+
"StackExchange|s0": true,
|
| 1038 |
+
"StackExchange|s1": true,
|
| 1039 |
+
"StackExchange|s2": true,
|
| 1040 |
+
"Github|s0": true,
|
| 1041 |
+
"Github|s1": true,
|
| 1042 |
+
"Github|s2": true
|
| 1043 |
+
},
|
| 1044 |
+
"fires": true
|
| 1045 |
+
},
|
| 1046 |
+
"C5_device_precision": {
|
| 1047 |
+
"required": "activations and operators on the same device at the same dtype",
|
| 1048 |
+
"per_model": {
|
| 1049 |
+
"1.4b": {
|
| 1050 |
+
"device": "cuda",
|
| 1051 |
+
"dtype": "float32",
|
| 1052 |
+
"gpu": "NVIDIA H100 NVL",
|
| 1053 |
+
"tf32": {
|
| 1054 |
+
"matmul_allow_tf32": false,
|
| 1055 |
+
"cudnn_allow_tf32": false,
|
| 1056 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 1057 |
+
}
|
| 1058 |
+
},
|
| 1059 |
+
"2.8b": {
|
| 1060 |
+
"device": "cuda",
|
| 1061 |
+
"dtype": "float32",
|
| 1062 |
+
"gpu": "NVIDIA H100 NVL",
|
| 1063 |
+
"tf32": {
|
| 1064 |
+
"matmul_allow_tf32": false,
|
| 1065 |
+
"cudnn_allow_tf32": false,
|
| 1066 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 1067 |
+
}
|
| 1068 |
+
}
|
| 1069 |
+
},
|
| 1070 |
+
"fires": true
|
| 1071 |
+
}
|
| 1072 |
+
},
|
| 1073 |
+
"PRIMARY": {
|
| 1074 |
+
"R_per_family_2.8b": {
|
| 1075 |
+
"multihop": 11.43874217650363,
|
| 1076 |
+
"multilingual": 14.505222414316746,
|
| 1077 |
+
"order-ops": 6.460960425429451,
|
| 1078 |
+
"poetry": 11.925362464834903,
|
| 1079 |
+
"typo": 11.414674645594424
|
| 1080 |
+
},
|
| 1081 |
+
"kendall_tau_vs_410m_2.8b": {
|
| 1082 |
+
"multihop": 0.6,
|
| 1083 |
+
"multilingual": 0.8,
|
| 1084 |
+
"order-ops": 0.8,
|
| 1085 |
+
"poetry": 0.6,
|
| 1086 |
+
"typo": 1.0
|
| 1087 |
+
},
|
| 1088 |
+
"n_families_R_ge_10": 4,
|
| 1089 |
+
"n_families_R_lt_5": 0,
|
| 1090 |
+
"n_families_tau_ge_0.6": 5
|
| 1091 |
+
},
|
| 1092 |
+
"VERDICT": "REPLICATES",
|
| 1093 |
+
"verdict_note": "the corpus effect is not an artifact of the low-competence regime at 410M",
|
| 1094 |
+
"declared_bias": [
|
| 1095 |
+
"Capability and scale are confounded; every larger model is also differently trained. No outcome licenses a causal reading \u2014 only 'the effect does / does not survive up the ladder'.",
|
| 1096 |
+
"The 5 corpora are a fixed panel, not a sample. n=5 is the replication unit for any claim about KINDS of corpus.",
|
| 1097 |
+
"N=25 here vs N=200 at 410M. S2 established reads are flat in N, so this is expected to be immaterial, but it is a difference and an attenuation must not be attributed to scale without checking N.",
|
| 1098 |
+
"Per-family seed SDs are thin: 3 seeds, 96-394 pairs per family. Thin denominators are what inflated the retired 58x figure. The SD is reported beside every ratio.",
|
| 1099 |
+
"The 410M reference is D3, which is EXPLORATORY and was computed BEFORE this pre-registration. The per-family estimand may not be described as chosen on capability grounds alone."
|
| 1100 |
+
],
|
| 1101 |
+
"provenance": {
|
| 1102 |
+
"schema_version": 1,
|
| 1103 |
+
"experiment": "CV6",
|
| 1104 |
+
"utc": "2026-08-30T06:31:18+00:00",
|
| 1105 |
+
"script": {
|
| 1106 |
+
"path": "experiments/cv6_per_family_ladder.py",
|
| 1107 |
+
"sha256": "f416c3cb55e95a02315504066479c2f2ba073c0d4676bbf09adb2f481cdb4a5b",
|
| 1108 |
+
"n_lines": 750
|
| 1109 |
+
},
|
| 1110 |
+
"git": {
|
| 1111 |
+
"commit": "c9c49697e295c0fec4225552836fc9733b279f75",
|
| 1112 |
+
"commit_short": "c9c49697e295",
|
| 1113 |
+
"git_repository_present": true,
|
| 1114 |
+
"branch": "pythia",
|
| 1115 |
+
"dirty": true,
|
| 1116 |
+
"dirty_files": [
|
| 1117 |
+
"CLAUDE.md",
|
| 1118 |
+
"README.md",
|
| 1119 |
+
"corpora/manifest_ood.json",
|
| 1120 |
+
"docs/HANDOFF.md",
|
| 1121 |
+
"docs/REPO_REVIEW_PROMPT.md",
|
| 1122 |
+
"docs/context/AGGREGATION_POLICY.md",
|
| 1123 |
+
"docs/context/CONTEXT.md",
|
| 1124 |
+
"docs/context/IMPLEMENTATION_RULES.md",
|
| 1125 |
+
"docs/context/STATE.md",
|
| 1126 |
+
"docs/experiments/INDEX.md",
|
| 1127 |
+
"docs/experiments/METRICS.md",
|
| 1128 |
+
"docs/experiments/PROPOSED.md",
|
| 1129 |
+
"docs/experiments/PROVENANCE.md",
|
| 1130 |
+
"docs/experiments/descriptions/E64.md",
|
| 1131 |
+
"docs/experiments/descriptions/EN1.md",
|
| 1132 |
+
"docs/experiments/preregs/CV3_margins.md",
|
| 1133 |
+
"docs/experiments/preregs/CV4_capability_ladder.md",
|
| 1134 |
+
"docs/experiments/preregs/CV5_metric_sensitivity.md",
|
| 1135 |
+
"docs/experiments/preregs/CV6_per_family_ladder.md",
|
| 1136 |
+
"docs/experiments/preregs/CV7_1b_rung.md",
|
| 1137 |
+
"docs/experiments/preregs/DN1_design_nulls.md",
|
| 1138 |
+
"docs/experiments/preregs/E33_logit_baseline.md",
|
| 1139 |
+
"docs/experiments/preregs/E36_qladder.md",
|
| 1140 |
+
"docs/experiments/preregs/E48_crossover.md",
|
| 1141 |
+
"docs/experiments/preregs/E48b_containment.md",
|
| 1142 |
+
"docs/experiments/preregs/E48c_exposure_vs_read.md",
|
| 1143 |
+
"docs/experiments/preregs/E51_interaction.md",
|
| 1144 |
+
"docs/experiments/preregs/E52_factorial.md",
|
| 1145 |
+
"docs/experiments/preregs/E53_ladder_summary.md",
|
| 1146 |
+
"docs/experiments/preregs/E55_matrix_robustness.md",
|
| 1147 |
+
"docs/experiments/preregs/E56_predictor_registry.md",
|
| 1148 |
+
"docs/experiments/preregs/E57_grid_variance_ci.md",
|
| 1149 |
+
"docs/experiments/preregs/E58_algebra_audit.md",
|
| 1150 |
+
"docs/experiments/preregs/E59_read_dose_response.md",
|
| 1151 |
+
"docs/experiments/preregs/E60_fitter_determinism.md",
|
| 1152 |
+
"docs/experiments/preregs/E61_randomized_network_null.md",
|
| 1153 |
+
"docs/experiments/preregs/E62_ladder1b_corrected_band.md",
|
| 1154 |
+
"docs/experiments/preregs/E65_training_axis_floor.md",
|
| 1155 |
+
"docs/experiments/preregs/E66b_determinism_floor.md",
|
| 1156 |
+
"docs/experiments/preregs/OS1_operator_space.md",
|
| 1157 |
+
"docs/experiments/preregs/P0_t52_parallel_port.md",
|
| 1158 |
+
"docs/experiments/preregs/R1_grid_rstrip.md",
|
| 1159 |
+
"docs/experiments/preregs/R2_e48_rstrip_arm.md",
|
| 1160 |
+
"docs/experiments/preregs/R3_ladder410_cpu.md",
|
| 1161 |
+
"docs/experiments/preregs/R4_corrections.md",
|
| 1162 |
+
"docs/experiments/preregs/R4b_e36_rstrip.md",
|
| 1163 |
+
"docs/experiments/preregs/R5_corpus_axis_uncertainty.md",
|
| 1164 |
+
"docs/experiments/preregs/R6_within_source_resampling.md",
|
| 1165 |
+
"docs/experiments/preregs/R8_ladder_rstrip_S2.md",
|
| 1166 |
+
"docs/experiments/preregs/R9_permutation_calibrated_min.md",
|
| 1167 |
+
"docs/experiments/preregs/superseded/PREREG_E48_CROSSOVER.md",
|
| 1168 |
+
"docs/reproducibility/ARTIFACTS.md",
|
| 1169 |
+
"docs/reproducibility/PAYLOAD_REDACTIONS.md",
|
| 1170 |
+
"docs/reproducibility/RELEASE.md",
|
| 1171 |
+
"docs/reproducibility/REVIEWER_REPRO.md",
|
| 1172 |
+
"docs/reproducibility/T1_FINDING_E33_BAND.md",
|
| 1173 |
+
"docs/reproducibility/T1_SWEEP_PREDICTIONS.md",
|
| 1174 |
+
"docs/reproducibility/T1_SWEEP_SPEC.md",
|
| 1175 |
+
"docs/reproducibility/review-snapshot-2026-08-19/MANIFEST.sha256",
|
| 1176 |
+
"docs/reproducibility/review-snapshot-2026-08-19/PROVENANCE.md",
|
| 1177 |
+
"docs/reproducibility/review-snapshot-2026-08-19/REDACTIONS.md",
|
| 1178 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/EXTERNAL_REVIEW.md",
|
| 1179 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/FINDINGS.md",
|
| 1180 |
+
"docs/reproducibility/review-snapshot-2026-08-19/review/reproducibility/PHASE_A_ACCOUNT.md"
|
| 1181 |
+
],
|
| 1182 |
+
"dirty_diff_sha256": "ec61dde1905616f407f51d5f6fe3ea37c9c6e3e1ade6998cd0644527d080c600",
|
| 1183 |
+
"describe": "c9c4969-dirty"
|
| 1184 |
+
},
|
| 1185 |
+
"argv": [
|
| 1186 |
+
"experiments/cv6_per_family_ladder.py",
|
| 1187 |
+
"--adjudicate"
|
| 1188 |
+
],
|
| 1189 |
+
"cwd": ".",
|
| 1190 |
+
"inputs": [
|
| 1191 |
+
{
|
| 1192 |
+
"path": "results/d3_corpus_by_family_410m.json",
|
| 1193 |
+
"sha256": "2ab2b53f90bd9d3435ebdb65aa0c9c1a82ac599a5e3c714f9dd9681d4bb9487c",
|
| 1194 |
+
"bytes": 23095,
|
| 1195 |
+
"exists": true
|
| 1196 |
+
},
|
| 1197 |
+
{
|
| 1198 |
+
"path": "results/cv6/cv6_1.4b_n25.json",
|
| 1199 |
+
"sha256": "4e8eab1ebedf0785a9e61db7fa77212cd2f89699fe627b08f7b9ea7b960cedae",
|
| 1200 |
+
"bytes": 30228,
|
| 1201 |
+
"exists": true
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"path": "results/cv6/cv6_2.8b_n25.json",
|
| 1205 |
+
"sha256": "7f1a6d31f1c8728c4c966d960e5502c6bfaaa8dddf4543cf5d9e5a8424f82b9f",
|
| 1206 |
+
"bytes": 32402,
|
| 1207 |
+
"exists": true
|
| 1208 |
+
},
|
| 1209 |
+
{
|
| 1210 |
+
"path": "results/cv6/cv6_c0_scorer_equivalence.json",
|
| 1211 |
+
"sha256": "2650f8b565cf128bc4ac01ee775849beec904a9a2329d4e75081492a4546d332",
|
| 1212 |
+
"bytes": 4013,
|
| 1213 |
+
"exists": true
|
| 1214 |
+
}
|
| 1215 |
+
],
|
| 1216 |
+
"env": {
|
| 1217 |
+
"python": "3.14.2",
|
| 1218 |
+
"torch": "2.13.0",
|
| 1219 |
+
"transformers": "5.14.1",
|
| 1220 |
+
"numpy": "2.5.1",
|
| 1221 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 1222 |
+
"machine": "arm64",
|
| 1223 |
+
"hostname": "<redacted for double-blind review>",
|
| 1224 |
+
"tf32": {
|
| 1225 |
+
"allow_tf32_matmul": false,
|
| 1226 |
+
"allow_tf32_cudnn": false,
|
| 1227 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 1228 |
+
},
|
| 1229 |
+
"threads": {
|
| 1230 |
+
"torch_num_threads": 10,
|
| 1231 |
+
"torch_num_interop_threads": 14,
|
| 1232 |
+
"OMP_NUM_THREADS": null,
|
| 1233 |
+
"os_cpu_count": 14
|
| 1234 |
+
}
|
| 1235 |
+
},
|
| 1236 |
+
"payload_sha256": "37c137de5d7788ff265c533db4dabc53540c7c9e6ad33e12b06d72182bbdd228"
|
| 1237 |
+
}
|
| 1238 |
+
}
|
results/cv7/cv7_1b_n200.json
ADDED
|
@@ -0,0 +1,821 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"experiment": "CV7 \u2014 the 1B rung, scoring only",
|
| 3 |
+
"prereg": "docs/experiments/preregs/CV7_1b_rung.md",
|
| 4 |
+
"status": "PRE-REGISTERED",
|
| 5 |
+
"decision_rule_verbatim": "Evaluated PER FAMILY; the verdict is the count across the five. Thresholds inherited VERBATIM from CV6_per_family_ladder.md, which fixed them before any ladder number existed. REPLICATES AT 1B if R(f,1B) >= 10 on >= 4 of 5 families AND ordering tau >= 0.6 on >= 3 of 5. ATTENUATES AT 1B if R(f,1B) < 5 on >= 4 of 5 families. UNCLEAR otherwise, including a split -- report the table and stop. Do not pool. Do not drop a family. Do not re-cut R across mismatched N.",
|
| 6 |
+
"model": "EleutherAI/pythia-1b-deduped",
|
| 7 |
+
"short": "1b",
|
| 8 |
+
"n_layers": 16,
|
| 9 |
+
"d_model": 2048,
|
| 10 |
+
"band": [
|
| 11 |
+
6,
|
| 12 |
+
7,
|
| 13 |
+
8,
|
| 14 |
+
9,
|
| 15 |
+
10,
|
| 16 |
+
11,
|
| 17 |
+
12,
|
| 18 |
+
13
|
| 19 |
+
],
|
| 20 |
+
"band_rule": "floor(0.38*L)..floor(0.92*L), layers < L-2",
|
| 21 |
+
"band_read_off_artifacts": {
|
| 22 |
+
"Wikipedia_en|s0": [
|
| 23 |
+
6,
|
| 24 |
+
7,
|
| 25 |
+
8,
|
| 26 |
+
9,
|
| 27 |
+
10,
|
| 28 |
+
11,
|
| 29 |
+
12,
|
| 30 |
+
13
|
| 31 |
+
],
|
| 32 |
+
"Wikipedia_en|s1": [
|
| 33 |
+
6,
|
| 34 |
+
7,
|
| 35 |
+
8,
|
| 36 |
+
9,
|
| 37 |
+
10,
|
| 38 |
+
11,
|
| 39 |
+
12,
|
| 40 |
+
13
|
| 41 |
+
],
|
| 42 |
+
"Wikipedia_en|s2": [
|
| 43 |
+
6,
|
| 44 |
+
7,
|
| 45 |
+
8,
|
| 46 |
+
9,
|
| 47 |
+
10,
|
| 48 |
+
11,
|
| 49 |
+
12,
|
| 50 |
+
13
|
| 51 |
+
],
|
| 52 |
+
"USPTO_Backgrounds|s0": [
|
| 53 |
+
6,
|
| 54 |
+
7,
|
| 55 |
+
8,
|
| 56 |
+
9,
|
| 57 |
+
10,
|
| 58 |
+
11,
|
| 59 |
+
12,
|
| 60 |
+
13
|
| 61 |
+
],
|
| 62 |
+
"USPTO_Backgrounds|s1": [
|
| 63 |
+
6,
|
| 64 |
+
7,
|
| 65 |
+
8,
|
| 66 |
+
9,
|
| 67 |
+
10,
|
| 68 |
+
11,
|
| 69 |
+
12,
|
| 70 |
+
13
|
| 71 |
+
],
|
| 72 |
+
"USPTO_Backgrounds|s2": [
|
| 73 |
+
6,
|
| 74 |
+
7,
|
| 75 |
+
8,
|
| 76 |
+
9,
|
| 77 |
+
10,
|
| 78 |
+
11,
|
| 79 |
+
12,
|
| 80 |
+
13
|
| 81 |
+
],
|
| 82 |
+
"Pile-CC|s0": [
|
| 83 |
+
6,
|
| 84 |
+
7,
|
| 85 |
+
8,
|
| 86 |
+
9,
|
| 87 |
+
10,
|
| 88 |
+
11,
|
| 89 |
+
12,
|
| 90 |
+
13
|
| 91 |
+
],
|
| 92 |
+
"Pile-CC|s1": [
|
| 93 |
+
6,
|
| 94 |
+
7,
|
| 95 |
+
8,
|
| 96 |
+
9,
|
| 97 |
+
10,
|
| 98 |
+
11,
|
| 99 |
+
12,
|
| 100 |
+
13
|
| 101 |
+
],
|
| 102 |
+
"Pile-CC|s2": [
|
| 103 |
+
6,
|
| 104 |
+
7,
|
| 105 |
+
8,
|
| 106 |
+
9,
|
| 107 |
+
10,
|
| 108 |
+
11,
|
| 109 |
+
12,
|
| 110 |
+
13
|
| 111 |
+
],
|
| 112 |
+
"StackExchange|s0": [
|
| 113 |
+
6,
|
| 114 |
+
7,
|
| 115 |
+
8,
|
| 116 |
+
9,
|
| 117 |
+
10,
|
| 118 |
+
11,
|
| 119 |
+
12,
|
| 120 |
+
13
|
| 121 |
+
],
|
| 122 |
+
"StackExchange|s1": [
|
| 123 |
+
6,
|
| 124 |
+
7,
|
| 125 |
+
8,
|
| 126 |
+
9,
|
| 127 |
+
10,
|
| 128 |
+
11,
|
| 129 |
+
12,
|
| 130 |
+
13
|
| 131 |
+
],
|
| 132 |
+
"StackExchange|s2": [
|
| 133 |
+
6,
|
| 134 |
+
7,
|
| 135 |
+
8,
|
| 136 |
+
9,
|
| 137 |
+
10,
|
| 138 |
+
11,
|
| 139 |
+
12,
|
| 140 |
+
13
|
| 141 |
+
],
|
| 142 |
+
"Github|s0": [
|
| 143 |
+
6,
|
| 144 |
+
7,
|
| 145 |
+
8,
|
| 146 |
+
9,
|
| 147 |
+
10,
|
| 148 |
+
11,
|
| 149 |
+
12,
|
| 150 |
+
13
|
| 151 |
+
],
|
| 152 |
+
"Github|s1": [
|
| 153 |
+
6,
|
| 154 |
+
7,
|
| 155 |
+
8,
|
| 156 |
+
9,
|
| 157 |
+
10,
|
| 158 |
+
11,
|
| 159 |
+
12,
|
| 160 |
+
13
|
| 161 |
+
],
|
| 162 |
+
"Github|s2": [
|
| 163 |
+
6,
|
| 164 |
+
7,
|
| 165 |
+
8,
|
| 166 |
+
9,
|
| 167 |
+
10,
|
| 168 |
+
11,
|
| 169 |
+
12,
|
| 170 |
+
13
|
| 171 |
+
]
|
| 172 |
+
},
|
| 173 |
+
"band_matches_rule": true,
|
| 174 |
+
"N": 200,
|
| 175 |
+
"n_prompts_per_arm": {
|
| 176 |
+
"Wikipedia_en|s0": 200,
|
| 177 |
+
"Wikipedia_en|s1": 200,
|
| 178 |
+
"Wikipedia_en|s2": 200,
|
| 179 |
+
"USPTO_Backgrounds|s0": 200,
|
| 180 |
+
"USPTO_Backgrounds|s1": 200,
|
| 181 |
+
"USPTO_Backgrounds|s2": 200,
|
| 182 |
+
"Pile-CC|s0": 200,
|
| 183 |
+
"Pile-CC|s1": 200,
|
| 184 |
+
"Pile-CC|s2": 200,
|
| 185 |
+
"StackExchange|s0": 200,
|
| 186 |
+
"StackExchange|s1": 200,
|
| 187 |
+
"StackExchange|s2": 200,
|
| 188 |
+
"Github|s0": 200,
|
| 189 |
+
"Github|s1": 200,
|
| 190 |
+
"Github|s2": 200
|
| 191 |
+
},
|
| 192 |
+
"operators_from": "results/ladder1b_b613/ (fitted for E62; NOT refitted here)",
|
| 193 |
+
"device": "cpu",
|
| 194 |
+
"dtype": "float32",
|
| 195 |
+
"operator_storage_dtype": "float16",
|
| 196 |
+
"device_caveat": "operators fitted on CUDA, scored on CPU. D2 measured a CUDA-vs-CPU cell-level divergence of 2.774e-04. This is D3's situation, which is what makes 1B-vs-410M the matched comparison.",
|
| 197 |
+
"torch": "2.13.0",
|
| 198 |
+
"K": [
|
| 199 |
+
1,
|
| 200 |
+
2,
|
| 201 |
+
5,
|
| 202 |
+
10,
|
| 203 |
+
20,
|
| 204 |
+
50,
|
| 205 |
+
100
|
| 206 |
+
],
|
| 207 |
+
"admitted_sets": [
|
| 208 |
+
"multihop",
|
| 209 |
+
"multilingual",
|
| 210 |
+
"order-ops",
|
| 211 |
+
"poetry",
|
| 212 |
+
"typo"
|
| 213 |
+
],
|
| 214 |
+
"corpora": [
|
| 215 |
+
"Wikipedia_en",
|
| 216 |
+
"USPTO_Backgrounds",
|
| 217 |
+
"Pile-CC",
|
| 218 |
+
"StackExchange",
|
| 219 |
+
"Github"
|
| 220 |
+
],
|
| 221 |
+
"seeds": [
|
| 222 |
+
0,
|
| 223 |
+
1,
|
| 224 |
+
2
|
| 225 |
+
],
|
| 226 |
+
"readout_convention": "STRIPPED (corrected), no prefix, flat-mean-7",
|
| 227 |
+
"pooling_convention": "flat pool over (item, intermediate) pairs WITHIN a family; rank = min over band then flat-mean-7 pass@k; z = max over band then mean",
|
| 228 |
+
"n_items": 449,
|
| 229 |
+
"n_pairs": 801,
|
| 230 |
+
"n_pairs_per_family": {
|
| 231 |
+
"multihop": 103,
|
| 232 |
+
"multilingual": 394,
|
| 233 |
+
"order-ops": 110,
|
| 234 |
+
"poetry": 98,
|
| 235 |
+
"typo": 96
|
| 236 |
+
},
|
| 237 |
+
"per_arm": {
|
| 238 |
+
"logit_I": {
|
| 239 |
+
"multihop": {
|
| 240 |
+
"rank": 0.24410540915395282,
|
| 241 |
+
"z": 3.898444828477878,
|
| 242 |
+
"n_pairs": 103
|
| 243 |
+
},
|
| 244 |
+
"multilingual": {
|
| 245 |
+
"rank": 0.1472081218274112,
|
| 246 |
+
"z": 2.7347642624287434,
|
| 247 |
+
"n_pairs": 394
|
| 248 |
+
},
|
| 249 |
+
"order-ops": {
|
| 250 |
+
"rank": 0.5207792207792208,
|
| 251 |
+
"z": 5.480567513812672,
|
| 252 |
+
"n_pairs": 110
|
| 253 |
+
},
|
| 254 |
+
"poetry": {
|
| 255 |
+
"rank": 0.0058309037900874635,
|
| 256 |
+
"z": 2.021673796736464,
|
| 257 |
+
"n_pairs": 98
|
| 258 |
+
},
|
| 259 |
+
"typo": {
|
| 260 |
+
"rank": 0.17708333333333331,
|
| 261 |
+
"z": 3.567758576323589,
|
| 262 |
+
"n_pairs": 96
|
| 263 |
+
}
|
| 264 |
+
},
|
| 265 |
+
"J|Wikipedia_en|s0": {
|
| 266 |
+
"multihop": {
|
| 267 |
+
"rank": 0.2538141470180305,
|
| 268 |
+
"z": 3.714084236367235,
|
| 269 |
+
"n_pairs": 103
|
| 270 |
+
},
|
| 271 |
+
"multilingual": {
|
| 272 |
+
"rank": 0.13741841914430747,
|
| 273 |
+
"z": 2.7458525141701102,
|
| 274 |
+
"n_pairs": 394
|
| 275 |
+
},
|
| 276 |
+
"order-ops": {
|
| 277 |
+
"rank": 0.4597402597402597,
|
| 278 |
+
"z": 5.137804371660406,
|
| 279 |
+
"n_pairs": 110
|
| 280 |
+
},
|
| 281 |
+
"poetry": {
|
| 282 |
+
"rank": 0.016034985422740525,
|
| 283 |
+
"z": 1.9996374888079507,
|
| 284 |
+
"n_pairs": 98
|
| 285 |
+
},
|
| 286 |
+
"typo": {
|
| 287 |
+
"rank": 0.3050595238095238,
|
| 288 |
+
"z": 3.7074354047266147,
|
| 289 |
+
"n_pairs": 96
|
| 290 |
+
}
|
| 291 |
+
},
|
| 292 |
+
"J|Wikipedia_en|s1": {
|
| 293 |
+
"multihop": {
|
| 294 |
+
"rank": 0.2538141470180305,
|
| 295 |
+
"z": 3.723332085655731,
|
| 296 |
+
"n_pairs": 103
|
| 297 |
+
},
|
| 298 |
+
"multilingual": {
|
| 299 |
+
"rank": 0.13778100072516317,
|
| 300 |
+
"z": 2.751266322687249,
|
| 301 |
+
"n_pairs": 394
|
| 302 |
+
},
|
| 303 |
+
"order-ops": {
|
| 304 |
+
"rank": 0.4584415584415584,
|
| 305 |
+
"z": 5.1602459907531735,
|
| 306 |
+
"n_pairs": 110
|
| 307 |
+
},
|
| 308 |
+
"poetry": {
|
| 309 |
+
"rank": 0.01749271137026239,
|
| 310 |
+
"z": 1.988615335858598,
|
| 311 |
+
"n_pairs": 98
|
| 312 |
+
},
|
| 313 |
+
"typo": {
|
| 314 |
+
"rank": 0.3050595238095238,
|
| 315 |
+
"z": 3.7324859214325747,
|
| 316 |
+
"n_pairs": 96
|
| 317 |
+
}
|
| 318 |
+
},
|
| 319 |
+
"J|Wikipedia_en|s2": {
|
| 320 |
+
"multihop": {
|
| 321 |
+
"rank": 0.2524271844660194,
|
| 322 |
+
"z": 3.706331000166032,
|
| 323 |
+
"n_pairs": 103
|
| 324 |
+
},
|
| 325 |
+
"multilingual": {
|
| 326 |
+
"rank": 0.13741841914430747,
|
| 327 |
+
"z": 2.7407319243846056,
|
| 328 |
+
"n_pairs": 394
|
| 329 |
+
},
|
| 330 |
+
"order-ops": {
|
| 331 |
+
"rank": 0.4610389610389611,
|
| 332 |
+
"z": 5.139317949251695,
|
| 333 |
+
"n_pairs": 110
|
| 334 |
+
},
|
| 335 |
+
"poetry": {
|
| 336 |
+
"rank": 0.016034985422740525,
|
| 337 |
+
"z": 1.9798154046340866,
|
| 338 |
+
"n_pairs": 98
|
| 339 |
+
},
|
| 340 |
+
"typo": {
|
| 341 |
+
"rank": 0.30357142857142855,
|
| 342 |
+
"z": 3.710389399745812,
|
| 343 |
+
"n_pairs": 96
|
| 344 |
+
}
|
| 345 |
+
},
|
| 346 |
+
"J|USPTO_Backgrounds|s0": {
|
| 347 |
+
"multihop": {
|
| 348 |
+
"rank": 0.29680998613037446,
|
| 349 |
+
"z": 4.109125902351824,
|
| 350 |
+
"n_pairs": 103
|
| 351 |
+
},
|
| 352 |
+
"multilingual": {
|
| 353 |
+
"rank": 0.15155910079767948,
|
| 354 |
+
"z": 2.907341442816964,
|
| 355 |
+
"n_pairs": 394
|
| 356 |
+
},
|
| 357 |
+
"order-ops": {
|
| 358 |
+
"rank": 0.48051948051948046,
|
| 359 |
+
"z": 5.347905703024431,
|
| 360 |
+
"n_pairs": 110
|
| 361 |
+
},
|
| 362 |
+
"poetry": {
|
| 363 |
+
"rank": 0.033527696793002916,
|
| 364 |
+
"z": 2.0518707047995863,
|
| 365 |
+
"n_pairs": 98
|
| 366 |
+
},
|
| 367 |
+
"typo": {
|
| 368 |
+
"rank": 0.4181547619047619,
|
| 369 |
+
"z": 4.076707179347674,
|
| 370 |
+
"n_pairs": 96
|
| 371 |
+
}
|
| 372 |
+
},
|
| 373 |
+
"J|USPTO_Backgrounds|s1": {
|
| 374 |
+
"multihop": {
|
| 375 |
+
"rank": 0.30513176144244103,
|
| 376 |
+
"z": 4.1576821566785425,
|
| 377 |
+
"n_pairs": 103
|
| 378 |
+
},
|
| 379 |
+
"multilingual": {
|
| 380 |
+
"rank": 0.1508339376359681,
|
| 381 |
+
"z": 2.907443309632995,
|
| 382 |
+
"n_pairs": 394
|
| 383 |
+
},
|
| 384 |
+
"order-ops": {
|
| 385 |
+
"rank": 0.4857142857142857,
|
| 386 |
+
"z": 5.373728732629256,
|
| 387 |
+
"n_pairs": 110
|
| 388 |
+
},
|
| 389 |
+
"poetry": {
|
| 390 |
+
"rank": 0.033527696793002916,
|
| 391 |
+
"z": 2.059163775085947,
|
| 392 |
+
"n_pairs": 98
|
| 393 |
+
},
|
| 394 |
+
"typo": {
|
| 395 |
+
"rank": 0.40625,
|
| 396 |
+
"z": 4.066875881205003,
|
| 397 |
+
"n_pairs": 96
|
| 398 |
+
}
|
| 399 |
+
},
|
| 400 |
+
"J|USPTO_Backgrounds|s2": {
|
| 401 |
+
"multihop": {
|
| 402 |
+
"rank": 0.30235783633841884,
|
| 403 |
+
"z": 4.154952114068188,
|
| 404 |
+
"n_pairs": 103
|
| 405 |
+
},
|
| 406 |
+
"multilingual": {
|
| 407 |
+
"rank": 0.15228426395939088,
|
| 408 |
+
"z": 2.9097203547655024,
|
| 409 |
+
"n_pairs": 394
|
| 410 |
+
},
|
| 411 |
+
"order-ops": {
|
| 412 |
+
"rank": 0.4753246753246753,
|
| 413 |
+
"z": 5.355981678312475,
|
| 414 |
+
"n_pairs": 110
|
| 415 |
+
},
|
| 416 |
+
"poetry": {
|
| 417 |
+
"rank": 0.03644314868804665,
|
| 418 |
+
"z": 2.0739537419713274,
|
| 419 |
+
"n_pairs": 98
|
| 420 |
+
},
|
| 421 |
+
"typo": {
|
| 422 |
+
"rank": 0.41517857142857145,
|
| 423 |
+
"z": 4.111865448455016,
|
| 424 |
+
"n_pairs": 96
|
| 425 |
+
}
|
| 426 |
+
},
|
| 427 |
+
"J|Pile-CC|s0": {
|
| 428 |
+
"multihop": {
|
| 429 |
+
"rank": 0.2787794729542302,
|
| 430 |
+
"z": 4.083725291548423,
|
| 431 |
+
"n_pairs": 103
|
| 432 |
+
},
|
| 433 |
+
"multilingual": {
|
| 434 |
+
"rank": 0.15445975344452503,
|
| 435 |
+
"z": 2.9522800377477316,
|
| 436 |
+
"n_pairs": 394
|
| 437 |
+
},
|
| 438 |
+
"order-ops": {
|
| 439 |
+
"rank": 0.43376623376623374,
|
| 440 |
+
"z": 5.306177348440344,
|
| 441 |
+
"n_pairs": 110
|
| 442 |
+
},
|
| 443 |
+
"poetry": {
|
| 444 |
+
"rank": 0.018950437317784254,
|
| 445 |
+
"z": 1.9803421716300809,
|
| 446 |
+
"n_pairs": 98
|
| 447 |
+
},
|
| 448 |
+
"typo": {
|
| 449 |
+
"rank": 0.39285714285714285,
|
| 450 |
+
"z": 4.1813286903003855,
|
| 451 |
+
"n_pairs": 96
|
| 452 |
+
}
|
| 453 |
+
},
|
| 454 |
+
"J|Pile-CC|s1": {
|
| 455 |
+
"multihop": {
|
| 456 |
+
"rank": 0.2815533980582524,
|
| 457 |
+
"z": 4.094569732841936,
|
| 458 |
+
"n_pairs": 103
|
| 459 |
+
},
|
| 460 |
+
"multilingual": {
|
| 461 |
+
"rank": 0.15772298767222623,
|
| 462 |
+
"z": 2.986847078749144,
|
| 463 |
+
"n_pairs": 394
|
| 464 |
+
},
|
| 465 |
+
"order-ops": {
|
| 466 |
+
"rank": 0.44025974025974024,
|
| 467 |
+
"z": 5.305445389314131,
|
| 468 |
+
"n_pairs": 110
|
| 469 |
+
},
|
| 470 |
+
"poetry": {
|
| 471 |
+
"rank": 0.018950437317784258,
|
| 472 |
+
"z": 2.0033713484905205,
|
| 473 |
+
"n_pairs": 98
|
| 474 |
+
},
|
| 475 |
+
"typo": {
|
| 476 |
+
"rank": 0.3988095238095238,
|
| 477 |
+
"z": 4.20682845885555,
|
| 478 |
+
"n_pairs": 96
|
| 479 |
+
}
|
| 480 |
+
},
|
| 481 |
+
"J|Pile-CC|s2": {
|
| 482 |
+
"multihop": {
|
| 483 |
+
"rank": 0.2787794729542302,
|
| 484 |
+
"z": 4.0395108611838335,
|
| 485 |
+
"n_pairs": 103
|
| 486 |
+
},
|
| 487 |
+
"multilingual": {
|
| 488 |
+
"rank": 0.1548223350253807,
|
| 489 |
+
"z": 2.943907567330631,
|
| 490 |
+
"n_pairs": 394
|
| 491 |
+
},
|
| 492 |
+
"order-ops": {
|
| 493 |
+
"rank": 0.43506493506493504,
|
| 494 |
+
"z": 5.280768312107433,
|
| 495 |
+
"n_pairs": 110
|
| 496 |
+
},
|
| 497 |
+
"poetry": {
|
| 498 |
+
"rank": 0.014577259475218656,
|
| 499 |
+
"z": 1.9708401384402294,
|
| 500 |
+
"n_pairs": 98
|
| 501 |
+
},
|
| 502 |
+
"typo": {
|
| 503 |
+
"rank": 0.3898809523809524,
|
| 504 |
+
"z": 4.157780651624004,
|
| 505 |
+
"n_pairs": 96
|
| 506 |
+
}
|
| 507 |
+
},
|
| 508 |
+
"J|StackExchange|s0": {
|
| 509 |
+
"multihop": {
|
| 510 |
+
"rank": 0.3370319001386962,
|
| 511 |
+
"z": 4.34045670449155,
|
| 512 |
+
"n_pairs": 103
|
| 513 |
+
},
|
| 514 |
+
"multilingual": {
|
| 515 |
+
"rank": 0.16243654822335027,
|
| 516 |
+
"z": 3.0495984307324826,
|
| 517 |
+
"n_pairs": 394
|
| 518 |
+
},
|
| 519 |
+
"order-ops": {
|
| 520 |
+
"rank": 0.4857142857142857,
|
| 521 |
+
"z": 5.51830342466181,
|
| 522 |
+
"n_pairs": 110
|
| 523 |
+
},
|
| 524 |
+
"poetry": {
|
| 525 |
+
"rank": 0.02040816326530612,
|
| 526 |
+
"z": 2.010183577938956,
|
| 527 |
+
"n_pairs": 98
|
| 528 |
+
},
|
| 529 |
+
"typo": {
|
| 530 |
+
"rank": 0.49107142857142855,
|
| 531 |
+
"z": 4.537338141351938,
|
| 532 |
+
"n_pairs": 96
|
| 533 |
+
}
|
| 534 |
+
},
|
| 535 |
+
"J|StackExchange|s1": {
|
| 536 |
+
"multihop": {
|
| 537 |
+
"rank": 0.334257975034674,
|
| 538 |
+
"z": 4.3061738963266025,
|
| 539 |
+
"n_pairs": 103
|
| 540 |
+
},
|
| 541 |
+
"multilingual": {
|
| 542 |
+
"rank": 0.16026105873821608,
|
| 543 |
+
"z": 3.0347819554662037,
|
| 544 |
+
"n_pairs": 394
|
| 545 |
+
},
|
| 546 |
+
"order-ops": {
|
| 547 |
+
"rank": 0.5012987012987012,
|
| 548 |
+
"z": 5.519914856824008,
|
| 549 |
+
"n_pairs": 110
|
| 550 |
+
},
|
| 551 |
+
"poetry": {
|
| 552 |
+
"rank": 0.021865889212827987,
|
| 553 |
+
"z": 2.0059123179134057,
|
| 554 |
+
"n_pairs": 98
|
| 555 |
+
},
|
| 556 |
+
"typo": {
|
| 557 |
+
"rank": 0.4672619047619047,
|
| 558 |
+
"z": 4.418834609289964,
|
| 559 |
+
"n_pairs": 96
|
| 560 |
+
}
|
| 561 |
+
},
|
| 562 |
+
"J|StackExchange|s2": {
|
| 563 |
+
"multihop": {
|
| 564 |
+
"rank": 0.334257975034674,
|
| 565 |
+
"z": 4.317730897838629,
|
| 566 |
+
"n_pairs": 103
|
| 567 |
+
},
|
| 568 |
+
"multilingual": {
|
| 569 |
+
"rank": 0.15917331399564902,
|
| 570 |
+
"z": 3.045205917266117,
|
| 571 |
+
"n_pairs": 394
|
| 572 |
+
},
|
| 573 |
+
"order-ops": {
|
| 574 |
+
"rank": 0.4987012987012987,
|
| 575 |
+
"z": 5.524625140970404,
|
| 576 |
+
"n_pairs": 110
|
| 577 |
+
},
|
| 578 |
+
"poetry": {
|
| 579 |
+
"rank": 0.02040816326530612,
|
| 580 |
+
"z": 2.000217623248392,
|
| 581 |
+
"n_pairs": 98
|
| 582 |
+
},
|
| 583 |
+
"typo": {
|
| 584 |
+
"rank": 0.4761904761904762,
|
| 585 |
+
"z": 4.479793511331081,
|
| 586 |
+
"n_pairs": 96
|
| 587 |
+
}
|
| 588 |
+
},
|
| 589 |
+
"J|Github|s0": {
|
| 590 |
+
"multihop": {
|
| 591 |
+
"rank": 0.2524271844660194,
|
| 592 |
+
"z": 3.8118835438802403,
|
| 593 |
+
"n_pairs": 103
|
| 594 |
+
},
|
| 595 |
+
"multilingual": {
|
| 596 |
+
"rank": 0.14213197969543148,
|
| 597 |
+
"z": 2.8611055270385775,
|
| 598 |
+
"n_pairs": 394
|
| 599 |
+
},
|
| 600 |
+
"order-ops": {
|
| 601 |
+
"rank": 0.4896103896103896,
|
| 602 |
+
"z": 5.030006443370472,
|
| 603 |
+
"n_pairs": 110
|
| 604 |
+
},
|
| 605 |
+
"poetry": {
|
| 606 |
+
"rank": 0.013119533527696792,
|
| 607 |
+
"z": 1.850219145303174,
|
| 608 |
+
"n_pairs": 98
|
| 609 |
+
},
|
| 610 |
+
"typo": {
|
| 611 |
+
"rank": 0.29910714285714285,
|
| 612 |
+
"z": 3.70514493167866,
|
| 613 |
+
"n_pairs": 96
|
| 614 |
+
}
|
| 615 |
+
},
|
| 616 |
+
"J|Github|s1": {
|
| 617 |
+
"multihop": {
|
| 618 |
+
"rank": 0.2621359223300971,
|
| 619 |
+
"z": 3.836685398828636,
|
| 620 |
+
"n_pairs": 103
|
| 621 |
+
},
|
| 622 |
+
"multilingual": {
|
| 623 |
+
"rank": 0.14648295866569977,
|
| 624 |
+
"z": 2.8760104446747543,
|
| 625 |
+
"n_pairs": 394
|
| 626 |
+
},
|
| 627 |
+
"order-ops": {
|
| 628 |
+
"rank": 0.5,
|
| 629 |
+
"z": 5.072230815887451,
|
| 630 |
+
"n_pairs": 110
|
| 631 |
+
},
|
| 632 |
+
"poetry": {
|
| 633 |
+
"rank": 0.014577259475218656,
|
| 634 |
+
"z": 1.8696253643352159,
|
| 635 |
+
"n_pairs": 98
|
| 636 |
+
},
|
| 637 |
+
"typo": {
|
| 638 |
+
"rank": 0.3080357142857143,
|
| 639 |
+
"z": 3.753457091632299,
|
| 640 |
+
"n_pairs": 96
|
| 641 |
+
}
|
| 642 |
+
},
|
| 643 |
+
"J|Github|s2": {
|
| 644 |
+
"multihop": {
|
| 645 |
+
"rank": 0.2579750346740638,
|
| 646 |
+
"z": 3.8195294026032234,
|
| 647 |
+
"n_pairs": 103
|
| 648 |
+
},
|
| 649 |
+
"multilingual": {
|
| 650 |
+
"rank": 0.145032632342277,
|
| 651 |
+
"z": 2.8609475317850785,
|
| 652 |
+
"n_pairs": 394
|
| 653 |
+
},
|
| 654 |
+
"order-ops": {
|
| 655 |
+
"rank": 0.4922077922077922,
|
| 656 |
+
"z": 5.031250395558097,
|
| 657 |
+
"n_pairs": 110
|
| 658 |
+
},
|
| 659 |
+
"poetry": {
|
| 660 |
+
"rank": 0.01020408163265306,
|
| 661 |
+
"z": 1.8547754169878912,
|
| 662 |
+
"n_pairs": 98
|
| 663 |
+
},
|
| 664 |
+
"typo": {
|
| 665 |
+
"rank": 0.30357142857142855,
|
| 666 |
+
"z": 3.727808784286026,
|
| 667 |
+
"n_pairs": 96
|
| 668 |
+
}
|
| 669 |
+
}
|
| 670 |
+
},
|
| 671 |
+
"logit_I_pooled_rank_flatmean7": 0.2190013977768011,
|
| 672 |
+
"provenance": {
|
| 673 |
+
"schema_version": 1,
|
| 674 |
+
"experiment": "CV7",
|
| 675 |
+
"utc": "2026-08-24T18:42:15+00:00",
|
| 676 |
+
"script": {
|
| 677 |
+
"path": "experiments/cv7_1b_rung.py",
|
| 678 |
+
"sha256": "17eed1b79d679e0044f836a0906667bbcfa1bcfe5dfe5399c99557321ea6f73d",
|
| 679 |
+
"n_lines": 316
|
| 680 |
+
},
|
| 681 |
+
"git": {
|
| 682 |
+
"commit": "180c1374d89790d6e60346e4d49581f5bba6b34f",
|
| 683 |
+
"commit_short": "180c1374d897",
|
| 684 |
+
"branch": "pythia",
|
| 685 |
+
"dirty": true,
|
| 686 |
+
"dirty_files": [
|
| 687 |
+
"experiments/cv7_1b_rung.py",
|
| 688 |
+
"figures/fA_effect_and_null.png",
|
| 689 |
+
"figures/fB_validity.png",
|
| 690 |
+
"figures/make_audit_figures.py",
|
| 691 |
+
"logs/",
|
| 692 |
+
"ocs/context/CONTEXT.md",
|
| 693 |
+
"paper/PAPER_HANDOFF.md",
|
| 694 |
+
"paper/audit_paper_5pp.tex",
|
| 695 |
+
"paper/audit_paper_7pp.tex",
|
| 696 |
+
"paper/build_audit.sh",
|
| 697 |
+
"paper/refs.bib",
|
| 698 |
+
"paper/repro-paper.tex",
|
| 699 |
+
"paper/tables_audit.tex",
|
| 700 |
+
"repro/30_repo_health.sh",
|
| 701 |
+
"repro/exp/cv7_1b.sh",
|
| 702 |
+
"repro/lib/banned_claims.py"
|
| 703 |
+
],
|
| 704 |
+
"dirty_diff_sha256": "d380ffb38dde5c7efcbe498d9454e038c0e71c01a846d485b439e379c597db29",
|
| 705 |
+
"describe": "180c137-dirty"
|
| 706 |
+
},
|
| 707 |
+
"argv": [
|
| 708 |
+
"experiments/cv7_1b_rung.py",
|
| 709 |
+
"--device",
|
| 710 |
+
"cpu"
|
| 711 |
+
],
|
| 712 |
+
"cwd": ".",
|
| 713 |
+
"inputs": [
|
| 714 |
+
{
|
| 715 |
+
"path": "results/ladder1b_b613/lens_Wikipedia_en_1b_n200_s0.pt",
|
| 716 |
+
"sha256": "f60c8ee8186e4d53e9a0195b5f0535506160d2d313b3bc174d3129b80d9426b2",
|
| 717 |
+
"bytes": 67112755,
|
| 718 |
+
"exists": true
|
| 719 |
+
},
|
| 720 |
+
{
|
| 721 |
+
"path": "results/ladder1b_b613/lens_Wikipedia_en_1b_n200_s1.pt",
|
| 722 |
+
"sha256": "25a07e77788297cc89a4603c0a1771d9cf2f8580a481c015bf77d039adcd0464",
|
| 723 |
+
"bytes": 67112755,
|
| 724 |
+
"exists": true
|
| 725 |
+
},
|
| 726 |
+
{
|
| 727 |
+
"path": "results/ladder1b_b613/lens_Wikipedia_en_1b_n200_s2.pt",
|
| 728 |
+
"sha256": "89838efaec6ce3b2f991ed142b320fd28c8f608e5205a8c934e0ac44675dd0fa",
|
| 729 |
+
"bytes": 67112755,
|
| 730 |
+
"exists": true
|
| 731 |
+
},
|
| 732 |
+
{
|
| 733 |
+
"path": "results/ladder1b_b613/lens_USPTO_Backgrounds_1b_n200_s0.pt",
|
| 734 |
+
"sha256": "4acda201f3fbc46f77e96d3f6f6840ab6cc53b8d80f694ec5080dab7321b054a",
|
| 735 |
+
"bytes": 67112825,
|
| 736 |
+
"exists": true
|
| 737 |
+
},
|
| 738 |
+
{
|
| 739 |
+
"path": "results/ladder1b_b613/lens_USPTO_Backgrounds_1b_n200_s1.pt",
|
| 740 |
+
"sha256": "4be9599eae66c8895479b2d0151fde7975929aedc420916b1b29d7415234eac1",
|
| 741 |
+
"bytes": 67112825,
|
| 742 |
+
"exists": true
|
| 743 |
+
},
|
| 744 |
+
{
|
| 745 |
+
"path": "results/ladder1b_b613/lens_USPTO_Backgrounds_1b_n200_s2.pt",
|
| 746 |
+
"sha256": "0f730745e88e5dae78b5f73b98a00ab55c28ed32dae8b7800ba8823445e090ff",
|
| 747 |
+
"bytes": 67112825,
|
| 748 |
+
"exists": true
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"path": "results/ladder1b_b613/lens_Pile-CC_1b_n200_s0.pt",
|
| 752 |
+
"sha256": "3bf4ce26087e107fa74da45922f6ecc486d6eab6fb29b7fbbf9b356a0ca4eabd",
|
| 753 |
+
"bytes": 67112685,
|
| 754 |
+
"exists": true
|
| 755 |
+
},
|
| 756 |
+
{
|
| 757 |
+
"path": "results/ladder1b_b613/lens_Pile-CC_1b_n200_s1.pt",
|
| 758 |
+
"sha256": "f41713739a1f278fd0d1c5b5c23db6ecfe1df0991945aa6b83ad716200e9a8f0",
|
| 759 |
+
"bytes": 67112685,
|
| 760 |
+
"exists": true
|
| 761 |
+
},
|
| 762 |
+
{
|
| 763 |
+
"path": "results/ladder1b_b613/lens_Pile-CC_1b_n200_s2.pt",
|
| 764 |
+
"sha256": "fe7a6a5b4579d5f47f0119d943efc700eb70220fb16acdec55c1f724dcde0ac2",
|
| 765 |
+
"bytes": 67112685,
|
| 766 |
+
"exists": true
|
| 767 |
+
},
|
| 768 |
+
{
|
| 769 |
+
"path": "results/ladder1b_b613/lens_StackExchange_1b_n200_s0.pt",
|
| 770 |
+
"sha256": "5048cd087976e256a7ff9f30d8066dd95d442229f8c2b0ff5f4d273eef512c5b",
|
| 771 |
+
"bytes": 67112769,
|
| 772 |
+
"exists": true
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"path": "results/ladder1b_b613/lens_StackExchange_1b_n200_s1.pt",
|
| 776 |
+
"sha256": "905a7e26d7af75d7e9a64a49ba5319156b0e220236c0bf8ec5c15d8989db2330",
|
| 777 |
+
"bytes": 67112769,
|
| 778 |
+
"exists": true
|
| 779 |
+
},
|
| 780 |
+
{
|
| 781 |
+
"path": "results/ladder1b_b613/lens_StackExchange_1b_n200_s2.pt",
|
| 782 |
+
"sha256": "78e10c625c615a13dc475fef35688ce8a90b8ab368ec84e1c8c7f88d70d3a040",
|
| 783 |
+
"bytes": 67112769,
|
| 784 |
+
"exists": true
|
| 785 |
+
},
|
| 786 |
+
{
|
| 787 |
+
"path": "results/ladder1b_b613/lens_Github_1b_n200_s0.pt",
|
| 788 |
+
"sha256": "459f997d6661ff07494e3ac0829864e02d71b15e0609875a705e032371a30158",
|
| 789 |
+
"bytes": 67112671,
|
| 790 |
+
"exists": true
|
| 791 |
+
},
|
| 792 |
+
{
|
| 793 |
+
"path": "results/ladder1b_b613/lens_Github_1b_n200_s1.pt",
|
| 794 |
+
"sha256": "187bd291fc66a6f3e5cf9a35c4640d9d963e1fd47434124dc3e7b7832f1a7950",
|
| 795 |
+
"bytes": 67112671,
|
| 796 |
+
"exists": true
|
| 797 |
+
},
|
| 798 |
+
{
|
| 799 |
+
"path": "results/ladder1b_b613/lens_Github_1b_n200_s2.pt",
|
| 800 |
+
"sha256": "e37d288145a6438be3c571ad3fac769b7fd314c00675573dc8e4e9b3812a2754",
|
| 801 |
+
"bytes": 67112671,
|
| 802 |
+
"exists": true
|
| 803 |
+
}
|
| 804 |
+
],
|
| 805 |
+
"env": {
|
| 806 |
+
"python": "3.14.2",
|
| 807 |
+
"torch": "2.13.0",
|
| 808 |
+
"transformers": "5.14.1",
|
| 809 |
+
"numpy": "2.5.1",
|
| 810 |
+
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
|
| 811 |
+
"machine": "arm64",
|
| 812 |
+
"hostname": "<redacted for double-blind review>",
|
| 813 |
+
"tf32": {
|
| 814 |
+
"allow_tf32_matmul": false,
|
| 815 |
+
"allow_tf32_cudnn": false,
|
| 816 |
+
"NVIDIA_TF32_OVERRIDE": "0"
|
| 817 |
+
}
|
| 818 |
+
},
|
| 819 |
+
"payload_sha256": "39a12fb962ca27e489684090b92af88ea55b4d2c0aab3e43061cd00eb5e627e0"
|
| 820 |
+
}
|
| 821 |
+
}
|