---
language:
- en
license: other
tags:
- security
- prompt-injection
- litert
---

# CLEF-Screen

Redo fine-tune of the CLEF-Flash `JointSchemaHead` (121,762,820 parameters,
122 tensors) on frozen Cloudflare CLEF-Flash LiteRT backbone hidden states.
Only the head is trained; the backbone is unmodified. This repository root
is a complete CLEF LiteRT checkpoint directory with the fine-tuned head.

## Training

- Compute: DigitalOcean GPU droplet `jev-retrain` (AMD Instinct MI355X, ROCm), encode 2026-11-04 11:31Z, head training finished 18:24Z.
- Method: `ft_clef_bench.py`, 3 epoch(s), lr 3e-05, batch 16,
  pos_weight 3.0. Training details and hashes: `redo-provenance.json`;
  full log: `redo-training-log.jsonl` when present.
- Head dtype in this repo: BF16 (deployment contract). Trained head was already BF16; no conversion.

## Evaluation status

Dev (21,977 rows) epoch0 AUC 0.7516 -> epoch3 AUC 0.9899, TPR@0.5 0.9911 FPR@0.5 0.2183, TPR@0.99 0.9583 FPR@0.99 0.0507. Script-run test banking+slack (21,965 rows) AUC 0.9897, TPR@0.99 0.9574 FPR@0.99 0.0537. Caveats: dev/test are 92% positive (~1,695 negatives); FPR@0.5 remains 0.22; test was evaluated by the training script at finish; no independent frozen replication yet. Full numbers: redo-final-results.json.

No frozen-test, adaptive-attack, or independent-replication result exists yet
for this checkpoint unless stated above. Treat all numbers as development
measurements, not deployment guarantees.