abacusai/Smaug-Mini

Dark-mode fix: theme-inherited text for benchmark names, brighter header blue

Move notes toggle below the LiveBench section

Fill Opus 4.6 AutomationBench (25.5) and JobBench (36.9) from announcement doc

Collapsible notes; lighten Training Approach; retitle and lighten Known Behaviors

Heat table: add NL2Repo-Bench row (Smaug-Mini 55.8)

Heat table LiveBench rows synced to published board scores

LiveBench category profile updated to published board scores

Remove unselected figure options

Finalize evaluation: keep multi-model heat table

MMMU-Pro comparison scores from Artificial Analysis

Option D: add Vision group with MMMU-Pro 75.6

HLE score 34.2

Option D: drop parenthetical qualifiers from eval names

Option D: drop Glimmer column, add LiveBench board scores for Qwen3.6-27B and Opus 4.6

Option D: Qwen3.7-Plus AutomationBench 20.4 (Artificial Analysis)

Option D: add HLE row (Smaug-Mini 33.8)

Option D: multi-model heat-shaded HTML table (comparison values from the Qwen3.8-27B card)

Citation: restore DPO-Positive entry

Citation: drop DPO-Positive entry

Citation: add base-model BibTeX (Qwen3.8, verbatim from the base card)

S5: drop run-level scores from behavioral table (they cite the transcribed replication run, not the S3 headline run)

Spell out per-benchmark base-score sources

Drop identical-protocols phrase from evaluation caption

Delta pill corner radius matched to bar chart

Remove superseded single vs-base figure

Three vs-base figure options (bars/dumbbell/table); radar delta bars aligned with table style

Radar axis labels clear of the circle

Trim outer canvas padding on figures (~5% larger render)

Delta bars restored with doc-matched geometry; doc-matched column and chart-table spacing

Agentic-style structure: model summary, delta values only

Delta bars: rectangles across centered zero-line, matching Flash

Move figure titles/captions into README text; larger in-figure text

Larger figure text: re-tune both charts for card display width, full-width embeds

Switch vs-base chart to dumbbell form

Drop redundant eval tables; regenerate vs-base chart with matching font scale

Replace LiveBench category table with the chart; re-render figure at 2x resolution

Add LiveBench category profile chart (radar + delta table)

Put Smaug-Mini as first column in all tables

Add vs-base benchmark figure, Abacus header, tags, citation

Add merged checkpoint-100 weights (vision + language)

Base model first in score tables; drop In-progress section

Add behavioral-differences section (paired AB trajectory analysis)

Remove LiveBench replication-methodology paragraph

Remove GPQA methodology paragraph

Drop model-card label from JobBench row (stated in note below)

Drop 3-run-average label from AutomationBench row

Simplify JobBench attribution, AB 3-run average, GPQA best-of-two, LiveBench best-of-runs note

Add Smaug-Mini model card (draft)

initial commit