Dataset Viewer
Auto-converted to Parquet Duplicate
task
stringclasses
7 values
type
stringclasses
1 value
mode
stringclasses
1 value
strategy
null
model
stringclasses
1 value
judge_model
stringclasses
1 value
inference_model
stringclasses
1 value
n_samples
int64
60
60
n_test_instances
int64
100
100
prompt
stringclasses
7 values
messages
stringclasses
7 values
raw_response
stringclasses
7 values
thinking
stringclasses
7 values
reasoning_summary
stringclasses
7 values
generated_hypotheses
stringclasses
7 values
known_hypotheses
stringclasses
6 values
usage
stringclasses
7 values
metrics
stringclasses
7 values
diversity
stringclasses
7 values
practical_utility
stringclasses
7 values
Deception Detection (deceptive_reviews)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou're a professional hotel review analyst.\nGiven a set of hotel reviews, we want to gen(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You're a professional hotel review analyst.\\nGiven a set of(...TRUNCATED)
"1. Truthful reviews are more likely to include concrete trip logistics—exact prices, dates, booki(...TRUNCATED)
"**Generating hypothesis pairs**\n\nI need to create ten hypothesis pairs based on reviews that pred(...TRUNCATED)
"**Generating hypothesis pairs**\n\nI need to create ten hypothesis pairs based on reviews that pred(...TRUNCATED)
"[\"Truthful reviews are more likely to include concrete trip logistics\\u2014exact prices, dates, b(...TRUNCATED)
"[\"**Spatial Configuration**: **Truthful opinions** emphasize **spatial configurations** (e.g., *sm(...TRUNCATED)
{"prompt_tokens": 11672, "completion_tokens": 5724, "reasoning_tokens": 5178}
"{\"per_hypothesis\": [{\"hypothesis\": \"Truthful reviews are more likely to include concrete trip (...TRUNCATED)
"{\"mean_pairwise_distance\": 0.29952528422529046, \"min_pairwise_distance\": 0.13060986995697021, \(...TRUNCATED)
"{\"accuracy\": 0.72, \"f1\": 0.7192003263973887, \"n_correct\": 72, \"n_total\": 100, \"n_other\": (...TRUNCATED)
Dreddit (dreaddit)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou're a psychologist and social scientist studying people's stress and their online post(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You're a psychologist and social scientist studying people's(...TRUNCATED)
"1. **Has stress:** Posts are more likely to indicate stress when the author uses urgent first-perso(...TRUNCATED)
"**Proposing hypothesis pairs**\n\nI need to provide 10 hypothesis pairs, balancing whether the stre(...TRUNCATED)
"**Proposing hypothesis pairs**\n\nI need to provide 10 hypothesis pairs, balancing whether the stre(...TRUNCATED)
"[\"**Has stress:** Posts are more likely to indicate stress when the author uses urgent first-perso(...TRUNCATED)
"[\"**Increased self-focus** in language, especially when expressing negative emotions, often indica(...TRUNCATED)
{"prompt_tokens": 6972, "completion_tokens": 3185, "reasoning_tokens": 2506}
"{\"per_hypothesis\": [{\"hypothesis\": \"**Has stress:** Posts are more likely to indicate stress w(...TRUNCATED)
"{\"mean_pairwise_distance\": 0.14107577403386434, \"min_pairwise_distance\": 0.07381284236907959, \(...TRUNCATED)
"{\"accuracy\": 0.76, \"f1\": 0.7519636213311286, \"n_correct\": 76, \"n_total\": 100, \"n_other\": (...TRUNCATED)
GPTGC (gptgc_detect)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou're a an AI generated content detection expert. You are great at detecting what type o(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You're a an AI generated content detection expert. You are g(...TRUNCATED)
"1. AI-generated text is more likely to open with broad atmospheric scene-setting and polished descr(...TRUNCATED)
"**Analyzing AI vs. Human Text Styles**\n\nI'm noticing that AI-generated text tends to start with p(...TRUNCATED)
"**Analyzing AI vs. Human Text Styles**\n\nI'm noticing that AI-generated text tends to start with p(...TRUNCATED)
"[\"AI-generated text is more likely to open with broad atmospheric scene-setting and polished descr(...TRUNCATED)
"[\"**AI-generated texts** exhibit token-level regularity and predictability, leading to lower perpl(...TRUNCATED)
{"prompt_tokens": 50928, "completion_tokens": 3063, "reasoning_tokens": 2643}
"{\"per_hypothesis\": [{\"hypothesis\": \"AI-generated text is more likely to open with broad atmosp(...TRUNCATED)
"{\"mean_pairwise_distance\": 0.19908152288860745, \"min_pairwise_distance\": 0.1254950761795044, \"(...TRUNCATED)
"{\"accuracy\": 0.81, \"f1\": 0.8089422006275939, \"n_correct\": 81, \"n_total\": 100, \"n_other\": (...TRUNCATED)
Headline Classification (headline_binary)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou are a professional writer for an online newspaper company. \nGiven a pair of headline(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a professional writer for an online newspaper compan(...TRUNCATED)
"1. Headlines that name a specific person, group, brand, or culprit and describe a concrete action t(...TRUNCATED)
"**Generating hypotheses for headlines**\n\nI need to create 10 hypotheses to explain why one headli(...TRUNCATED)
"**Generating hypotheses for headlines**\n\nI need to create 10 hypotheses to explain why one headli(...TRUNCATED)
"[\"Headlines that name a specific person, group, brand, or culprit and describe a concrete action t(...TRUNCATED)
"[\"Headlines that evoke negative emotions or emotional intensity are more likely to attract clicks,(...TRUNCATED)
{"prompt_tokens": 4009, "completion_tokens": 4296, "reasoning_tokens": 3922}
"{\"per_hypothesis\": [{\"hypothesis\": \"Headlines that name a specific person, group, brand, or cu(...TRUNCATED)
"{\"mean_pairwise_distance\": 0.41951173345247905, \"min_pairwise_distance\": 0.17964214086532593, \(...TRUNCATED)
"{\"accuracy\": 0.52, \"f1\": 0.5198079231692677, \"n_correct\": 52, \"n_total\": 100, \"n_other\": (...TRUNCATED)
GPTGC (llamagc_detect)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou're a an AI generated content detection expert. You are great at detecting what type o(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You're a an AI generated content detection expert. You are g(...TRUNCATED)
"1. AI-generated texts are more likely to begin with a clean orientation block that quickly names th(...TRUNCATED)
"**Planning hypothesis pairs**\n\nI need to create 10 hypothesis pairs that are balanced across outc(...TRUNCATED)
"**Planning hypothesis pairs**\n\nI need to create 10 hypothesis pairs that are balanced across outc(...TRUNCATED)
"[\"AI-generated texts are more likely to begin with a clean orientation block that quickly names th(...TRUNCATED)
"[\"**AI-generated texts** exhibit token-level regularity and predictability, leading to lower perpl(...TRUNCATED)
{"prompt_tokens": 45004, "completion_tokens": 2526, "reasoning_tokens": 2070}
"{\"per_hypothesis\": [{\"hypothesis\": \"AI-generated texts are more likely to begin with a clean o(...TRUNCATED)
"{\"mean_pairwise_distance\": 0.2623837232589722, \"min_pairwise_distance\": 0.18334943056106567, \"(...TRUNCATED)
"{\"accuracy\": 0.65, \"f1\": 0.6178622120318811, \"n_correct\": 65, \"n_total\": 100, \"n_other\": (...TRUNCATED)
Persuasive Arguments (persuasive_pairs)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou are an intelligent rhetorician and debater who masters persuasiveness in language.\nG(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are an intelligent rhetorician and debater who masters p(...TRUNCATED)
"1. The more persuasive version often sounds more certain and decisive, using firm claims like “wi(...TRUNCATED)
"**Considering persuasive arguments**\n\nI think I should come up with a balance of hypotheses expla(...TRUNCATED)
"**Considering persuasive arguments**\n\nI think I should come up with a balance of hypotheses expla(...TRUNCATED)
"[\"The more persuasive version often sounds more certain and decisive, using firm claims like \\u20(...TRUNCATED)
"[\"**Two-sided arguments** tend to be more persuasive for highly educated people as proponents put (...TRUNCATED)
{"prompt_tokens": 4948, "completion_tokens": 6207, "reasoning_tokens": 5696}
"{\"per_hypothesis\": [{\"hypothesis\": \"The more persuasive version often sounds more certain and (...TRUNCATED)
"{\"mean_pairwise_distance\": 0.4396168781651391, \"min_pairwise_distance\": 0.2300865650177002, \"m(...TRUNCATED)
"{\"accuracy\": 0.91, \"f1\": 0.909918927034331, \"n_correct\": 91, \"n_total\": 100, \"n_other\": 0(...TRUNCATED)
Retweet Prediction (retweet)
real
few_shot
null
gpt-5.4-2026-03-05
gpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18
60
100
"[SYSTEM]\nYou are a social media expert. You are an expert at determining which tweet will be retwe(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a social media expert. You are an expert at determin(...TRUNCATED)
"1. The first tweet is more likely to be retweeted when it expresses a clear takeaway, insight, or v(...TRUNCATED)
"**Creating balanced hypotheses**\n\nI’m considering how to structure hypotheses for tweets. Since(...TRUNCATED)
"**Creating balanced hypotheses**\n\nI’m considering how to structure hypotheses for tweets. Since(...TRUNCATED)
"[\"The first tweet is more likely to be retweeted when it expresses a clear takeaway, insight, or v(...TRUNCATED)
"[\"Tweets that express strong sentiments, especially negative emotions, are more likely to be retwe(...TRUNCATED)
{"prompt_tokens": 5453, "completion_tokens": 3622, "reasoning_tokens": 3177}
"{\"per_hypothesis\": [{\"hypothesis\": \"The first tweet is more likely to be retweeted when it exp(...TRUNCATED)
"{\"mean_pairwise_distance\": 0.26803329424424605, \"min_pairwise_distance\": 0.12073051929473877, \(...TRUNCATED)
"{\"accuracy\": 0.57, \"f1\": 0.562608076492727, \"n_correct\": 57, \"n_total\": 100, \"n_other\": 0(...TRUNCATED)

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

hypobench-few_shot-gpt5420260305-20260415_013728

Benchmark: hypobench Generated: 2026-04-15T01:53:21.164092

Parameters

  • model: gpt-5.4-2026-03-05
  • judge_model: gpt-4o-mini-2024-07-18
  • inference_model: gpt-4o-mini-2024-07-18
  • mode: few_shot
  • strategy: none
  • n_hypotheses: 10
  • n_samples: 60
  • n_test_instances: 100
  • n_tasks: 7
  • avg_clarity: 3.403 +/- 0.382
  • avg_novelty: 3.006 +/- 0.299
  • avg_plausibility: 4.022 +/- 0.159
  • avg_quality: 3.477 +/- 0.17
  • avg_diversity: 0.29 +/- 0.109
  • avg_accuracy: 0.706 +/- 0.136
  • avg_f1: 0.699 +/- 0.14

Results (Real Tasks)

Task Clarity Novelty Plausibility Quality Diversity Accuracy F1
Deception Detection (deceptive_reviews) 3.8 3.1 3.9 3.595 0.300 0.720 0.719
Dreddit (dreaddit) 3.1 3.1 4.1 3.433 0.141 0.760 0.752
GPTGC (gptgc_detect) 3.1 3.2 4.0 3.433 0.199 0.810 0.809
Headline Classification (headline_binary) 3.8 2.9 4.2 3.633 0.420 0.520 0.520
GPTGC (llamagc_detect) 3.2 3.4 3.8 3.459 0.262 0.650 0.618
Persuasive Arguments (persuasive_pairs) 3.8 2.9 4.2 3.633 0.440 0.910 0.910
Retweet Prediction (retweet) 3.0 2.5 4.0 3.152 0.268 0.570 0.563

Run Command

benchmarks/hypobench/run.py --model gpt-5.4-2026-03-05 --judge-model gpt-4o-mini-2024-07-18 --inference-model gpt-4o-mini-2024-07-18 --mode few_shot --n-hypotheses 10 --n-samples 60 --n-test-instances 100 --reasoning-effort high --upload-hf
Downloads last month
29