YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

mixSWE Qwen3.8-27B prefix-KV artifacts

Trained prefix-KV exports for the mixSWE 4-segment (4x1024) task, Qwen3.8-27B. Each epoch_NN/artifact.tgz is a serving-ready export (FA K/V + re-captured GDN states, 128 files + CAPTURE manifest). See RUN_REGISTRATION.json per arm for full provenance (rows sha, hyperparams, code shas).

Arm Rows Loss Steps Final
seg4v3 (BFCL, CE+KL) 852 CE 1.0 CE + 0.3 KL(fwd) 648 (54x12) epoch_12, loss 0.01
seg4uld (BFCL, CE+KL+UL) 852 CE + 226 UL (pre-shifted) 1.0 CE + 0.3 KL(rev) + 0.1 UL 804 (67x12) epoch_12, loss 0.03

Training code: main@927ddead-era xorl (fp32 fix, KL/UL disjoint masks, int64), FA-only (GDN frozen: mixswe harness offsets are not 64-aligned), DP2/FSDP2.

Per-epoch exports (01-11) come from retrains with save_epochs=1 (the original runs saved only the final DCP); epoch_12 here is the original final and will be replaced by the retrain's epoch_12 for a consistent set (originals kept under *_original_final/).

Deviations (read before using)

  • seg2v3 (swepro rows) trained the multi-text span (row/span cross from 1-index confusion; multi was planned for the swepro span symmetrically). Its artifact is dynamics evidence only.
  • v1 geometries are NOT 64-aligned (DSH harness 1252 % 64 = 36): seg2v3/seg4v3/ seg4uld artifacts (n_inject 4324/4239) are refused by the stock Q27-hybrid connector, which requires n_inject % 64 == 0. v2 geometries pad the harness to alignment (DSH +28 -> 1280, BFCL +49 -> 3264) and fix the row/span mapping; all v2 retrains supersede these artifacts.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support