Laya NLI memory-conflict head β v4 & three-run protocol
Collection
5 items β’ Updated
laya-nli-conflict-v9 β research archive (round 9), NOT delivered
β οΈ Research archive β NOT a delivered model. This checkpoint failed its round's acceptance gates and was never shipped. The current production head is
slow-stack/laya-nli-memory-conflict(v4). Uploaded 2026-10-02 for provenance/backup while round 11 (multi-run verdict protocol) waits for Kaggle GPU quota.
Round 9 (2026-09-30) added +400 rows on the v8 corpus (byte-identical carried blocks): attribute top-up 200 (including 79/120 B2-isomorphic shapes), attribute reverse-control 40 (kind β true), waver 120, change-already-happened 40.
Headline results (frozen 1000-pair main val unless noted)
- main val 0.9050 β (program best), val_soft 0/300 β (first zero since v4), in-distribution probe 30/30, conformal re-cut dual-rule pass (selection budget 0.34 locked into the tooling)
- failed: B2 0.7491 β 0.889 β doubling the isomorphic dose made it worse (dose-response falsified at this dosage; round 10 later showed single-run B2 readings are dominated by run variance), old-20 18/20 (β, thin-edge flips)
Artifacts
| file | value |
|---|---|
| model.safetensors | SHA256 885f256fβ¦e606d6 (full hash in archive_sha256_manifest.txt) |
| rl_agent_config.json | Ο(noul) = 1.1233; encoder jhu-clsp/mmBERT-base; bf16 |
| metrics.json | val_accuracy 0.905, val_ece 0.0201, n_val 1000, no_rl true |
| val_probs.json | frozen-val probability dump (calibration analyses) |
Provenance
- Training: Kaggle GPU kernel
daphnelaurent/laya-nli-conflict-cev5, datasetdaphnelaurent/nli-conflict-pairsv14 - Round record & full gate table:
kaggle_eval/HANDOFF_NLI_V9.md - Base: fine-tuned from convaiinnovations/laya-multilingual lineage; bf16 weights;
checkpoint_latest/intentionally not uploaded
Model tree for slow-stack/laya-nli-conflict-v9
Base model
convaiinnovations/laya-multilingual