Datasets:

silence-suzuki
/

FIRE-Bench-verified

task_id large_stringlengths 10 36	research_question large_stringlengths 0 327	instruction large_stringlengths 875 3.9k	instruction_gt large_stringlengths 0 21.8k	conclusion large_stringlengths 118 373	dataset_source large_stringclasses 9 values	has_local_data bool 2 classes
activation_control	Can we efficiently elicit long chain-of-thought reasoning in language models through activation-level interventions?	You are a research agent. Conduct research and experiment about the question: "Can we efficiently elicit long chain-of-thought reasoning in language models through activation-level interventions?" You have access to the following resources: Models: - Load with HuggingFace: Qwen/Qwen2.5-7B - Load with HuggingFace: Qwe...	You are a research agent. Conduct research and experiment about the question: "Can we efficiently elicit long chain-of-thought reasoning in language models through activation-level interventions?" You have access to the following resources: Models: - Load with HuggingFace: Qwen/Qwen2.5-7B - Load with HuggingFace: Qwe...	Efficiently eliciting long chain-of-thought reasoning in language models is possible through activation-level interventions. By amplifying a small set of high-impact activations in the last few layers and inserting wait tokens, long-form reasoning can be invoked without additional training, enhancing self-reflection ra...		false
awareness_detection	Can frontier language models detect that a given interaction transcript comes from an evaluation rather than real-world deployment?	You are a research agent. Conduct research and experiment about the question: "Can frontier language models detect that a given interaction transcript comes from an evaluation rather than real-world deployment?" Note: Different wording of prompts can lead to different experiment results. You have access to the follow...	You are a research agent. Conduct research and experiment about the question: "Can frontier language models detect that a given interaction transcript comes from an evaluation rather than real-world deployment?" Note: Different wording of prompts can lead to different experiment results. You have access to the follow...	Frontier language models can detect that a given interaction transcript comes from an evaluation rather than real-world deployment with significant accuracy.	Download dataset.json and put it in /data folder https://huggingface.co/datasets/jjpn2/eval_awareness	true
charxiv_chart_understanding	Can multimodal LLMs accurately understand and reason about real-world scientific charts, as opposed to simplified template charts?	You are a research agent. Conduct research and experiment about the question: "Can multimodal LLMs accurately understand and reason about real-world scientific charts, as opposed to simplified template charts?" You have access to the following resources: Models (test ALL of them): - OpenAI's gpt-4o - OpenAI's gpt-4o-...		Multimodal LLMs struggle to accurately understand and reason about real-world scientific charts, achieving less than half accuracy on reasoning questions and performing significantly worse than humans.		false
cot_faithfulness_gaps	To what extent do reasoning models’ chains-of-thought faithfully reflect their internal reasoning processes when they exploit external hints?	You are a research agent. Conduct research and experiment about the question: "To what extent do reasoning models’ chains-of-thought faithfully reflect their internal reasoning processes when they exploit external hints?" You have access to the following resources: Models: - Claude's claude-3-7-sonnet-20250219, claude...	You are a research agent. Conduct research and experiment about the question: "To what extent do reasoning models’ chains-of-thought faithfully reflect their internal reasoning processes when they exploit external hints?" You have access to the following resources: Models: - Claude's claude-3-7-sonnet-20250219, claude...	Reasoning models' chains-of-thought rarely reflect their internal reasoning processes when exploiting external hints.		true
cot_in_planning	Does chain-of-thought (CoT) prompting truly enable large language models (LLMs) to learn generalizable algorithmic reasoning abilities, or does it merely rely on highly specific, pattern-matching prompts?	You are a research agent. Conduct research and experiment about the question: "Does chain-of-thought (CoT) prompting truly enable large language models (LLMs) to learn generalizable algorithmic reasoning abilities, or does it merely rely on highly specific, pattern-matching prompts?" You have access to the following r...	You are a research agent. Conduct research and experiment about the question: "Does chain-of-thought (CoT) prompting truly enable large language models (LLMs) to learn generalizable algorithmic reasoning abilities, or does it merely rely on highly specific, pattern-matching prompts?" You have access to the following r...	Chain-of-thought prompting does not reliably enable large language models to learn generalizable algorithmic reasoning abilities. It relies on highly specific, pattern-matching prompts, and performance degrades with increased problem complexity.		false
cot_without_prompting	Can large language models, without any chain of thought prompts, reveal reasoning paths and improve answer accuracy by altering its decoding approach?	You are a research agent. Conduct research and experiment about the question: "Can large language models, without any chain of thought prompts, reveal reasoning paths and improve answer accuracy by altering its decoding approach?" You have access to the following resources: Model: - mistralai/Mistral-7B-v0.1 - Huggin...	You are a research agent. Conduct research and experiment about the question: "Can large language models, without any chain of thought prompts, reveal reasoning paths and improve answer accuracy by altering its decoding approach?" You have access to the following resources: Model: - mistralai/Mistral-7B-v0.1 - Huggin...	Large language models can reveal reasoning paths and improve answer accuracy by altering the decoding approach. Exploring alternative token sequences uncovers hidden reasoning trajectories, and selecting the path with the highest answer confidence significantly outperforms standard greedy decoding.		true
counterfactual_simulatability	Do natural language explanations provided by language models enable humans to accurately simulate the model's behavior under counterfactual inputs?	You are a research agent. Conduct research and experiment about the question: "Do natural language explanations provided by language models enable humans to accurately simulate the model's behavior under counterfactual inputs?" You have access to the following resources: Models: - gpt-3.5-turbo and gpt-4 via the prov...	You are a research agent. Conduct research and experiment about the question: "Do natural language explanations provided by language models enable humans to accurately simulate the model's behavior under counterfactual inputs?" You have access to the following resources: Models: - gpt-3.5-turbo and gpt-4 via the prov...	Natural language explanations provided by language models do not enable humans to accurately simulate the model's behavior under counterfactual inputs.		false
distributive_fairness	How fair are large language models when making resource allocation decisions across different demographic groups?	You are a research agent. Conduct research and experiment about the question: "How fair are large language models when making resource allocation decisions across different demographic groups?" You have access to the following resources: Models: - GPT-4o, GPT-4o-mini - Claude's claude-3-5-sonnet-20241022 - You can ca...	You are a research agent. Conduct research and experiment about the question: "How fair are large language models when making resource allocation decisions across different demographic groups?" You have access to the following resources: Models: - GPT-4o, GPT-4o-mini - Claude's claude-3-5-sonnet-20241022 - You can ca...	Large language models are poorly aligned with human distributional fairness preferences in resource allocation decisions across different demographic groups. They struggle to use resources like money to reduce inequality and are sensitive to prompt and template changes, though they perform better when selecting from pr...		false
fallback_behaviors	What behaviors do language models exhibit when they are uncertain, and how do these fallback patterns manifest across different models and tasks?	You are a research agent. Conduct research and experiment about the question: "What behaviors do language models exhibit when they are uncertain, and how do these fallback patterns manifest across different models and tasks?" You have access to the following resources: Models: - Load with HuggingFace: meta-llama/Llam...	You are a research agent. Conduct research and experiment about the question: "What behaviors do language models exhibit when they are uncertain, and how do these fallback patterns manifest across different models and tasks?" You have access to the following resources: Models: - Load with HuggingFace: meta-llama/Llam...	Language models exhibit a consistent ordering of fallback behaviors under uncertainty, shifting from sequence repetitions to degenerate text to hallucinations as they become more advanced. This pattern also appears within single-generation trajectories as uncertainty increases.		false
fractal_complexity_of_language	When and why do LLMs deviate from the narrow fractal parameter range characteristic of natural language, as visualized by Holder and Hurst exponents?	You are a research agent. Conduct research and experiment about the question: "When and why do LLMs deviate from the narrow fractal parameter range characteristic of natural language, as visualized by Holder and Hurst exponents?" You have access to the following resources: Models: - Openai's gpt-3.5-turbo - Gemini's g...	You are a research agent. Conduct research and experiment about the question: "When and why do LLMs deviate from the narrow fractal parameter range characteristic of natural language, as visualized by Holder and Hurst exponents?" You have access to the following resources: Models: - Openai's gpt-3.5-turbo - Gemini's g...	LLMs deviate from the narrow fractal parameter range characteristic of natural language when factors like decoding temperature and prompting method are adjusted, even if log-perplexity scores remain stable.	Pick a small subset of GAGLE dataset named processed_data.jsonl in /data/gagle folder https://huggingface.co/datasets/ibomohsin/gagle	true

End of preview. Expand in Data Studio

FIRE-Bench (verified)

A benchmark of 35 hand-curated research tasks from the FIRE-Bench project. Unlike the auto-generated companion dataset silence-suzuki/FIRE-Bench-unverified, these have been written and reviewed manually -- prompts, ground-truth plans, and conclusions are all human-validated.

Schema

field	description
`task_id`	unique identifier (e.g. `activation_control`)
`research_question`	the question the agent must answer
`instruction`	full prompt the agent sees (research question + resources)
`instruction_gt`	ground-truth procedural plan (used for evaluation, not shown to the agent)
`conclusion`	ground-truth answer; what the agent's final write-up is compared against
`dataset_source`	upstream URL or short note for the data, when the curators left one
`has_local_data`	true when raw data files are bundled at `tasks/<task_id>/data/` in this repo

Usage

from datasets import load_dataset
ds = load_dataset("silence-suzuki/FIRE-Bench-verified", split="train")

for task in ds:
    output = my_agent.run(task["instruction"])
    # evaluate output against task["conclusion"]

The raw per-task files (instruction.txt, instruction_gt.txt, conclusion.txt) are also available under tasks/<task_id>/ for filesystem-walking workflows. For tasks where the curators bundled local data, tasks/<task_id>/data/ holds the dataset files (JSONL, JSON, images, etc.) and tasks/<task_id>/dataset.txt documents the source.

To pull just the assets for one task:

from huggingface_hub import snapshot_download
snapshot_download(
    "silence-suzuki/FIRE-Bench-verified",
    repo_type="dataset",
    allow_patterns=["tasks/lost_in_the_middle/*", "tasks/lost_in_the_middle/**"],
)

For the runner / scoring code see maitrix-org/FIRE-Bench.

Downloads last month: 462

Number of rows:

Total file size:

1.64 GB