Skip to content

Latest commit

 

History

History
61 lines (51 loc) · 3.67 KB

File metadata and controls

61 lines (51 loc) · 3.67 KB

Binary vs. multi-class classification benchmark

Model: gemini-3.7-flash | Scales: [1000, 3000] | Seeds: [17, 29, 43] | Classification sample size: 40

Classification accuracy

Not run this pass (--skip-classification).

Dirty-data recovery (local only, no LLM)

"Rows" is the actual balanced training size used, which can be smaller than the scale header: _max_balanced_count caps every request at smallest_class_size × num_classes so a stratified sample never asks a rare class for more rows than it has (see trec_coarse below, whose rarest coarse class has only 86 examples -- both scale requests land on the same capped, and therefore identical, result).

Scale 1000

Dataset Kind Rows Dirty raw acc Dirty optimized acc Recovery Clean delta
yelp_polarity binary 1000 0.7840 0.8387 +0.0547 +0.0000
amazon_polarity binary 1000 0.7133 0.7660 +0.0527 +0.0000
imdb binary 1000 0.7433 0.8147 +0.0713 +0.0000
rotten_tomatoes binary 1000 0.6467 0.6513 +0.0047 +0.0000
sst2 binary 1000 0.6540 0.6480 -0.0060 +0.0000
subj binary 1000 0.7840 0.8300 +0.0460 +0.0000
tweet_eval_irony binary 1000 0.5827 0.6160 +0.0333 +0.0000
tweet_eval_hate binary 1000 0.5353 0.5220 -0.0133 +0.0047
cr binary 1000 0.7132 0.7451 +0.0319 +0.0000
amazon_counterfactual binary 1000 0.7873 0.8173 +0.0300 +0.0000
ag_news multi 1000 0.6873 0.7193 +0.0320 +0.0000
dbpedia_14 multi 1000 0.7940 0.7733 -0.0207 +0.0000
yahoo_answers_topics multi 1000 0.4373 0.4153 -0.0220 +0.0000
emotion multi 1000 0.4175 0.4015 -0.0160 +0.0000
tweet_eval_emotion multi 1000 0.4390 0.4289 -0.0102 +0.0000
tweet_eval_sentiment multi 1000 0.4407 0.4440 +0.0033 +0.0000
20_newsgroups multi 1000 0.3173 0.3247 +0.0073 -0.0087
trec_coarse multi 516 † 0.6975 0.6728 -0.0247 +0.0000
tweet_sentiment_extraction multi 1000 0.5120 0.5140 +0.0020 +0.0000

Scale 3000

Dataset Kind Rows Dirty raw acc Dirty optimized acc Recovery Clean delta
yelp_polarity binary 3000 0.8027 0.8727 +0.0700 +0.0000
amazon_polarity binary 3000 0.7773 0.8233 +0.0460 +0.0000
imdb binary 3000 0.7680 0.8607 +0.0927 +0.0047
rotten_tomatoes binary 3000 0.6713 0.7013 +0.0300 +0.0000
sst2 binary 3000 0.7020 0.7520 +0.0500 -0.0020
subj binary 3000 0.8260 0.9067 +0.0807 +0.0000
tweet_eval_irony binary 2834 † 0.5787 0.6053 +0.0267 +0.0000
tweet_eval_hate binary 3000 0.5813 0.5647 -0.0167 -0.0093
cr binary 2458 † 0.7145 0.7708 +0.0564 +0.0110
amazon_counterfactual binary 1054 † 0.7633 0.8120 +0.0487 +0.0000
ag_news multi 3000 0.7420 0.8120 +0.0700 +0.0000
dbpedia_14 multi 3000 0.8233 0.9173 +0.0940 +0.0000
yahoo_answers_topics multi 3000 0.5040 0.5500 +0.0460 +0.0000
emotion multi 3000 0.5976 0.6961 +0.0985 -0.0118
tweet_eval_emotion multi 1176 † 0.4797 0.4661 -0.0136 -0.0257
tweet_eval_sentiment multi 3000 0.4627 0.4687 +0.0060 +0.0000
20_newsgroups multi 3000 0.4647 0.5373 +0.0727 -0.0127
trec_coarse multi 516 † 0.6975 0.6728 -0.0247 +0.0000
tweet_sentiment_extraction multi 3000 0.5480 0.6073 +0.0593 +0.0027

† capped below the requested scale by the smallest class's available rows.