Fredred89/Predator-Q-Backup-2026-06-23

Predator-Q Backup (2026-06-23)

This repo contains a backup of all Predator-Q artifacts from the 2026-06-23 session.

Contents

• imatrix/qwopus_imatrix.gguf — Importance matrix for IQ-quants (14 MB, 100 chunks WikiText-2)
• moq_recipes/*.txt — All 9 MoQ tensor recipes from kaitchup HF (3.0-5.0 BPW)
• kl_div_base_logits/base_logits_0.dat — F16 reference logits for KL-div evaluation (5.9 GB)
• wikitext-2-raw/wiki.test.raw — WikiText-2 test set (1.3 MB)
• eval_logs/kl_div_*.log — Per-build KL-divergence evaluation logs
• lcb30_results/lcb_*.json — LCB-30 full results per build
• lcb30_results/lcb_*.log — LCB-30 raw logs
• build_logs/build_*.log — Quantize build logs

Companion GGUF

The actual GGUF file (MoQ-4.0 winner, 12.6 GB) lives at: https://huggingface.co/Fredred89/Qwopus3.6-27B-Coder-GGUF-Predator-Q

Reproducing

To regenerate from this backup:

# 1. Download imatrix + base model
# (qimx) and qwopus_f16.gguf from elsewhere (not backed up due to size)

# 2. Apply any recipe from moq_recipes/
llama-quantize --imatrix qwopus_imatrix.gguf \
  --tensor-type-file moq_recipes/MoQ_qwen3.6-27b_tensors_4.0.txt \
  qwopus_f16.gguf output.gguf Q4_K

# 3. Evaluate KL-div using base_logits_0.dat
llama-perplexity -m output.gguf -f wikitext-2-raw/wiki.test.raw \
  -c 512 -b 512 -np 1 \
  --kl-divergence-base kl_div_base_logits/base_logits_0.dat \
  --kl-divergence --seed 1337 -ngl 99 --chunks 50

Phase Status (2026-06-23)

• Phase 1: MoQ recipe study + base logits gen — DONE
• Phase 2: Hand-tune initial recipes — DONE (used kaitchup's directly)
• Phase 3: ASI-Evolve recipe search — IN PROGRESS
• Phase 4: Final validation + HF upload — DONE (winner uploaded)

Winner (Phase 1+2)

MoQ-4.0 with imatrix: 12.6 GB, LCB-30 = 50.0% (15/30), KL-div = 0.034

• 18% smaller than Q4_K_M (15.4 GB)
• 26% smaller than 4A BEST (16.98 GB)
• Same LCB-30 pass rate
• KL-div is 0.034 (vs Q4_K_M 0.019 — within Pareto frontier)