Qwen3.6-27B-FP8 full microbench N=5 — think vs no-think (clean FP8 redo)#34
Merged
Lightheartdevs merged 4 commits intoJun 1, 2026
Merged
Conversation
The clean FP8 redo of the 27B run the 397B entry had to exclude as a Q8/FP8 serving failure — 112 cells, 0 runaways, 0 HTTP-400 storms. Headline: thinking is net-negative for 27B on this bench (no-think 35/60 = 58%, think trends lower). Smoking gun: p2_triage 0/5 think vs 5/5 no-think (overthinks a closed label set). Thinking's only win is p3_business (length discipline, not insight). Qualitative: 27B is the content winner on p3 longform (best bias recall, sharpest synthesis, best citation honesty) held back by a word-count grader artifact and market scrape-loop variance. In-progress: no-think complete (60/60); think p3 tail still landing. Hand-grading dimensions not yet filled. Will finalize when the grid completes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…fixes) Grid complete (119/120 graded). Corrects the WIP draft's quantitative errors that the audit caught: - finish-reason reality: 113/119 clean done_signal, 6 errors (5 HTTP-400 + 1 max_tokens) — NOT "0 errors / 112 clean". The 6 are model loop-into-ctx- overflow on market/business-think, not FP8 instability (the honest framing of "FP8 viable where Q8 failed"). - final aggregates: no-think 35/60, think 29/60 (think net-negative, complete). - p3_market-think corrected to 3/4 (61 iters / 32k tok), v3 dropped-stuck. - p3_doc word-cap stated as real 705-vs-700 near-miss, not invented 697/711. - p3_writing think 1/5 vs no-think 5/5 added (think breaks it). Adds the registration the repo contract requires: manifest.json (provenance + finish-reason census + run inventory), 3 claims.yaml entries, index rows in hardware-tests/README.md and root README.md, and market_v3.skip-reason. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lightheartdevs
changed the title
…or, taxonomy note Tier-1 audit fixes: state temp=0.3 and the 500W cap (vs 397B's 600W) in the reproduce block; clarify think p3_market is 3/4 (v3 dropped) so the arm is 29/59 graded; add the microbench-index 'where this lives' taxonomy note. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…p8-microbench-2026-05-31 # Conflicts: # claims.yaml
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The clean FP8 redo of the Qwen3.6-27B run that the 397B entry had to exclude as a Q8/FP8 serving failure. Full MMBT 12-family microbench at N=5, both reasoning modes. Grid complete: 120 cells attempted, 119 graded.
FP8 serving is stable (the gap the 397B entry flagged is closed)
done_signalThe 6 FP8 errors are the model looping itself into the 131072 context ceiling on the two hardest tasks (market, business-think — e.g. business-think_v2 errored at iter 186 / 12M cumulative prompt-tokens), not quant instability.
Headline — thinking is net-negative for 27B
No-think 35/60 (58%) vs think 29/60 (48%).
p2_triage0/5 think vs 5/5 no-think (reasons off the closed label set).p3_writingalso breaks under thinking (1/5 vs 5/5).p3_business(5/5 vs 1/5) — length discipline, not insight.Qualitative
27B is the content winner on p3 longform (best bias recall 8/8 — beats 397B-think's 5/8; sharpest synthesis; best citation honesty) held back by a near-miss word cap (briefs ~700 by
wc, grader counts 703–705) and market scrape-loop variance.Registration (repo contract)
manifest.json— provenance, image digest, serving config, finish-reason census, run inventoryclaims.yamlentries (serving-viable, thinking-net-negative, triage-overthinking)hardware-tests/README.md+ rootREADME.mdmarket_v3.skip-reasonfor the dropped operator-killed repKnown open items (not blockers)
p3_docword-count grader counts markdown tokens — worth a grader fix before quoting doc pass-rates.🤖 Generated with Claude Code