Vec2Seq reviewer checkpoints for Qwen2.5-0.5B-Instruct
This repository contains a deliberately compact reviewer subset: 7
selected Vec2Seq attacker checkpoints (614291928 bytes). It does not
contain the frozen victim model or tokenizer. Obtain Qwen/Qwen2.5-0.5B-Instruct at immutable
revision 7ae557604adf67be50417f59c2c2f167def9a775 under its upstream terms.
The subset fixes training seed 1234 to cover the three primary observation
surfaces, full-sequence reconstruction at width 32, the direct-only prefix mode,
and both corrected alternative poolers. Seed 1234 is a coverage choice, not an
independent statistical sample. The companion code artifact contains the complete
632-row evidence registry and statistics; this repository supports the strict
two-dataset audit listed in metadata/reviewer_evaluations.csv.
Each runs/<slug>/ directory contains one checkpoint and its sanitized resolved
configuration. The raw .pt files preserve the exact production SHA-256 values.
They contain attacker-module state and runtime metadata but no victim weights,
optimizer state, or EMA payload. Verify SHA256SUMS before loading them.
Use the frozen protocol named by each manifest row. Primary and length rows use
cache_mask_v1; alternative-pooler and direct-only rows use
cache_mask_modes_v1 with both architecture fields supplied explicitly. The
validity mask, locked dataset hashes, metric definitions, and reference outputs
are specified in the companion code artifact.
The other selected checkpoints are formally deferred. Consequently this subset does not by itself regenerate every multi-seed interval, table, or figure.