Data for 'Semantic Anchors or Random Codes? A Two-Domain Diagnostic of Text-Conditioned Behavior Pretraining'
收藏资源简介：
# Data for "Semantic Anchors or Random Codes? A Two-Domain Diagnostic of Text-Conditioned Behavior Pretraining" This record contains the per-run experimental outputs underlying every quantitative claim in the paper. Every file is checksummed in `RELEASE_MANIFEST.json` (SHA-256, generated 2026-09-13, revised 2026-10-25 to 380 files: the revision adds 6 M4 representation-isolation feature npz). ## Contents Data are stored as **one ZIP archive per cohort**; inner paths inside each archive match the relpaths listed in `RELEASE_MANIFEST.json` exactly (e.g., `main_results_v3/m4base_text_s1_idrec_v2.json`). Verification: unzip an archive and hash each inner file against the manifest. Archives larger than 2 GB are **split into 2 GB parts** (`.001`, `.002`, …) for upload reliability. Reassemble by concatenating the parts in order (`copy /b main_results_v2.zip.001+main_results_v2.zip.002+… main_results_v2.zip` on Windows, `cat main_results_v2.zip.* > main_results_v2.zip` on Linux), or open the `.001` file directly with 7-Zip; then unzip and hash inner files against the manifest. `PARTS_MD5.txt` lists the MD5 checksum of every file in this record (parts and whole files) as stored. | Archive / file | Files | Role in the paper | |---|---|---| | `main_results_v3.zip` | 126 | Walker v3 corrected-reanalysis per-run JSONs (leave-one-out centroids, aligned windows, literal Top-2, permutation-tail test). **The claim source for all Walker numbers** (Table I, Figs. 2–4). | | `main_results_v2.zip`, `main_results.zip` | 130 + 58 | Walker legacy archives (prespecified-evaluator generation). Archived for provenance; **not** the claim source. | | `v2_aggregates.zip` | 5 | Aggregate lineage v2/v3/v4 (paper Table I). | | `cheetah_c5.zip` | 39 | Cheetah B5 cohort: 34 per-run JSONs + 5 audit feature archives (npz). | | `cheetah_c4.zip` | 22 | Cheetah C4 representation-isolation addendum: 15 per-run JSONs + `aggregate_c4.json` + 6 feature npz (M4 protocol). | | `RELEASE_MANIFEST.json` | — | Per-file SHA-256 checksums for all 380 data files. | | `FINAL_PROPOSAL.md` | 1 | The internally archived pre-data protocol, published **byte-identical** to the archived version (SHA-256 `52a19afd…`, re-verified 2026-10-25) — original formatting blemishes are intentionally unmodified, as evidence of no post-hoc editing. | | — (M5 addendum) | — | The M5 permuted-target control (post-outcome, exploratory; endpoints G1–G4) is documented in the paper (Section V, Table III) and in the header comment of `experiments/run_permuted.sh` in the code repository. The internal protocol-revision working file is not part of this release. | ## Code Training, injection, evaluation (v3 evaluator, hard-coded reset seeds) and aggregation code: **https://github.com/life9036/semantic-anchors-diagnostic** — branch `exp/anchors`, tags `anchors-v1` (frozen main-matrix state) and `anchors-v2` (adds the M5 permuted-target control). Checkouts/snapshots (83 GB) are not redistributed here; per-run JSONs in this record are sufficient to re-derive every table and figure number. ## Notes - Walker results use the v3 evaluator — a post-hoc corrected reanalysis of the prespecified evaluator; corrected endpoints are treated as exploratory in the paper (see abstract and Methods). - The paper's Data Availability section mirrors the mapping above. ## License CC-BY-4.0 (this record). Code: MIT (see the GitHub repository).
