Skip to content

GlossoBench leaderboard — ms

Status: Live · Audience: All · bench_version: 1.0

Real --intensity max ("fairest") and fast-intensity runs on the ms language config. Judge-free (no paid judge required). Use the filter bar to narrow by board / language / intensity / run kind, or search by model name. Click any column header to sort — the MS column sorts descending by default; ranks re-number after sort + filter.

  • Fairest board: 6 models, full 5-axis breakdown.
  • Fast board: 10 runs, composite MS only (cheap intensity).
  • Source: docs/leaderboard/ms/v1/results_v1.jsonl (one JSON object per run, each carries source_commit).

How to add a result: append a row to the JSONL, then re-run python website/scripts/render_leaderboard.py (or mkdocs build). See docs/leaderboard/FILEGUIDE.md for the file format.

No rows match the current filters.

Fairest board (--intensity max)

Full 5-axis (Knowledge / NLU / IF / NLG / Cultural). MS = register-composite mean score across the judge-free axes. Safety* proxy not shown here (separate axis; see README axis table).

#ModelMSKnowledgeNLUIFNLGCulturalJudgeWall (s)MeasuredCommit
1SEA-LION v4.5 27B-IT
v45
81.779.593.776.965.193.1judge-free1198.92026-07-253101af6
2Qwen3.5-27B (v4.5 base lineage)
qwen35-27b
80.380.994.073.164.089.7judge-free1171.32026-07-253101af6
3Qwen3.6-27B
qwen36-27b
77.574.991.273.165.582.8judge-free1290.62026-07-253101af6
4Gemma-4-31B-IT
gemma
76.583.193.673.139.893.1judge-free1796.72026-07-253101af6
5SEA-LION-v3-8B
sealion8b
69.758.780.161.562.186.2judge-free416.12026-07-253101af6
6SEA-LION-E2B-IT-2B
e2b
55.451.073.253.830.269.0judge-free1049.92026-07-253101af6

Fast board (--intensity fast)

Cheap-intensity composite MS + wall time. fast_seeds3 = 3-seed stability re-run of the same fast flags.

ModelMSWall (s)Run kindMeasuredCommit
SEA-LION v4.5 27B-IT
v45
80.5172.5fast2026-07-253101af6
SEA-LION v4.5 27B-IT
v45
80.5157.5fast_seeds32026-07-253101af6
Qwen3.5-27B (v4.5 base lineage)
qwen35-27b
78.5148.8fast2026-07-253101af6
Gemma-4-31B-IT
gemma
71.9277.6fast_seeds32026-07-253101af6
Gemma-4-31B-IT
gemma
71.9299.2fast2026-07-253101af6
Qwen3.6-27B
qwen36-27b
68.4165.5fast2026-07-253101af6
SEA-LION-v3-8B
sealion8b
67.899.9fast2026-07-253101af6
SEA-LION-v3-8B
sealion8b
67.887.9fast_seeds32026-07-253101af6
SEA-LION-E2B-IT-2B
e2b
53.8205.0fast2026-07-253101af6
SEA-LION-E2B-IT-2B
e2b
53.8193.5fast_seeds32026-07-253101af6

Backend / methodology parity checks

These used different flags than the main board — kept separate so they are never averaged into a ranking. Not directly comparable to the boards above.

backend_parity_check — SEA-LION v4.5 27B-IT

  • Flags: --thinking on (matched flags across backends)
  • Backends: hf: MS 25.2 · vllm: MS 26.0
  • Verdict: within CI -> hf and vllm backends equivalent after the v1.1.0 vLLM fixes (prompt_token_ids tokenization, skip_special_tokens, think-strip parity verified identical)
  • Note: NOT comparable to the fairest board above -- --thinking on burns the response budget very differently than the fairest board's flags, so this is a much lower absolute MS by construction. A backend-equivalence check, not a ranking entry.
  • Commit: 3101af6 · Measured: 2026-07-25

Provenance: every row carries a source_commit and measured_at date. Scores are immutable per bench_version; old rows stay valid forever.

A separate CI-generated nightly demo board (small fixed model roster, directional Elo, explicitly NOT a GlossoBench-7 claim) is at leaderboard/ms/v1/index.html.

Related: README (GitHub) · METHODOLOGY.md · leaderboard file guide