laya-nli-conflict-v5-ce — research archive (round 5, arm B), NOT delivered

⚠️ Research archive — 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 5 (2026-09-28) arm B: pure CE on graded soft targets, same corpus as arm A (see laya-nli-conflict-v5).

Headline results (frozen 1000-pair main val unless noted)

• main val 0.902 ✓ (over v4's 0.901), new-10 10/10 ✓
• failed: val_soft 7 errors (✗), polarity +2 (✗)
• Key finding: graded soft targets widened the confidence band 0.28 → 16.2pp in both arms (the compression-band main attack succeeded; arm B's main-val bins recovered monotonicity). Combined with arm A's collapse, this established pure CE as the round-6+ main arm.

Artifacts

file	value
model.safetensors	SHA256 997bd810…db8717 (full hash in archive_sha256_manifest.txt)
rl_agent_config.json	τ(noul) = 1.0647; encoder jhu-clsp/mmBERT-base; bf16
metrics.json	val_accuracy 0.902, val_ece 0.0261, n_val 1000, no_rl true
val_probs.json	frozen-val probability dump (calibration analyses)

Provenance

• Training: Kaggle GPU kernel daphnelaurent/laya-nli-conflict-ce v1, dataset daphnelaurent/nli-conflict-pairs v10
• Round record & full gate table: kaggle_eval/HANDOFF_NLI_V5.md
• Base: fine-tuned from convaiinnovations/laya-multilingual lineage; bf16 weights; checkpoint_latest/ intentionally not uploaded