Discriminative diverse-text rescue test (Mistral-7B 4K, FP16-failing operating point, n=40)

[jagmarques]

Summary

• Adds experiments/kaggle/nq_mistral_diverse_rescue/ (2 files): kernel .py + kernel-metadata.json
• Discriminative NIAH harness using completion prompt on wikitext haystack (no chat template) -- the operating point where FP16 partially fails at 4K context
• Tests FP16, K2V2_pb0, K3V2_pb0 on Mistral-7B-Instruct-v0.3, n=40 trials, ctx=4096
• Primary model: NousResearch/Mistral-7B-Instruct-v0.3 (ungated mirror, no token needed)
• Quant code (quantize_kv / get_kv / set_kv / to_dyn / device handling) is byte-identical to nq_yi_rescue.py (proven reference); random.Random(SEED) reseeded per config for label-independent needle pairing
• Outputs non_discriminative error key if FP16 hits 0 or 40 (saturation/degeneration detected)

Proof of work

VERIFY-WITH: kaggle kernels push -p experiments/kaggle/nq_mistral_diverse_rescue

OUTPUT:
Kernel push error: Maximum weekly GPU quota of 30.00 hours reached.

BLOCKED-KAGGLE: GPU quota exhausted (30h/week ceiling hit by earlier kernels today).
Files are built, ast-parsed (AST OK), secret-scanned (SCANNER: gitleaks - clean, exit 0).
Kernel slug: jagmarques/nq-mistral-diverse-rescue (ready to push when quota resets Monday).

AST check: python3.11 -c "import ast; ast.parse(open('nq_mistral_diverse_rescue.py').read())" -> AST OK
Secret scan: SCANNER: gitleaks - clean, exit 0

Diff stat:
 experiments/kaggle/nq_mistral_diverse_rescue/kernel-metadata.json    |  20 +
 experiments/kaggle/nq_mistral_diverse_rescue/nq_mistral_diverse_rescue.py | 466 +++++++++++++++++++++
 2 files changed, 486 insertions(+)