kv-compress: PRGF artifacts (Qwen3-8B, RULER-4K, KVPress protocol)

Code: github.com/triquang26/RestoreKV, branch claude/modal-gpu-restorekv-setup-v4gszh (all numbers and their derivation: docs/prgf/RESULTS.md). Paths mirror the Modal volume prgf-runs (mounted at /runs), so /runs/<path> in the code is <path> here.

Headline (all 6500 RULER-4K rows, paired with the leaderboard predictions)

Contents

• ckpt/prgf_v4plus/final - main model. Eval spec: {"method": "prgf", "adapter": "<dir>", "slots_per_region": 2, "num_global": 2, "plus": true}
• ckpt/prgf_v1, prgf_v3, prgf_v4 - lineage (v1: 7+1 slots; v3: + region reconstruction; v4: 16 slots)
• ckpt/abl_{A,B,C,D}/final - controlled 2x2 ablation from RestoreKV+ (A causal, B partition, C causal+recon, D partition+recon; use "mask_mode": "causal" for A and C)
• ckpt/v5_ctrl, ckpt/v5_read - read-matching pilot (negative result)
• data/train.jsonl - 2700 contexts / ~11.5k teacher QA pairs (LongAlpaca, PG-19, Tulu-3 FLAN; Qwen3-8B answers); data/contexts_*.jsonl raw contexts; data/kvzip_scores, data/kvzip_plus_scores cached per-context scores (training speed-up only); data/query_moment.safetensors (transport ablation)
• results/<run>/<split>/<ratio>.json (+ _preds.jsonl) - every evaluation with per-sample predictions; results/merged/ - the assembled 6500-row prediction sets for 16x and 20x
• leaderboard/ - KVPress leaderboard per-sample predictions (16x) used for the paired comparisons
• analysis/ - scripts that produced the tables (paired, task-stratified bootstrap)

Reproduce on a new machine

git clone -b claude/modal-gpu-restorekv-setup-v4gszh https://github.com/triquang26/RestoreKV && cd RestoreKV
hf sync hf://buckets/twanghcmut/kv-compress ./runs      # this bucket
# kvpress pinned to commit 66c5b0903b3be742bcba6b96d9796bd561e58f2d (see kv_image in modal_prgf.py)