Skip to content

Commit 6506ad0

Browse files
unamedkrclaude
andcommitted
bench: improve Document-Level RAG test prompts + Q4_K dequant investigation
Findings from Karpathy loop: 1. Q4_K_M GGUF on-the-fly dequant (TQ_NO_Q4) has a bug — garbage output. Root cause: likely in fused_dot_q4_k or Metal dispatch. Does NOT affect normal users (auto Q4 conversion is the default and works correctly). 2. Auto Q4 path (Q8_0 → Q4) works for fact extraction but with artifacts: "Santos" → "SanSannt", "Park" → "PPar". Semantically correct, but character-level precision is lost. 3. 1B model instruction-following is the real bottleneck for Document QA. The model enters "Step 1: Identify..." reasoning chains instead of directly answering. Better prompts help but don't solve it fully. 4. GGUF on-the-fly Q4_K bug confirmed: logits have -2.68 bias vs Q8_0's +3.98. Scale/min handling in fused_dot_q4_k suspected but unit test shows correct formula. Bug may be in per-tensor type dispatch for mixed Q4_K/Q6_K models (Q4_K_M uses both). Action: Document QA concept is valid, needs 3B+ Q8 models for reliable demonstration. The Q4_K on-the-fly bug is tracked for future fix. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1 parent 57a2918 commit 6506ad0

206 files changed

Lines changed: 30742 additions & 0 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

.gitignore

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -63,3 +63,6 @@ tq_run.dSYM/
6363
docs/assets/hero_backup.png
6464
build_nomt/
6565
build_nometal/
66+
67+
.env
68+

0 commit comments

Comments
 (0)