EconEvals/EconEvalsJune2026

Dataset Card for EconEvals

EconEvals is an evaluation suite for measuring language model capabilities on work performed in the US labor economy, organized around the Department of Labor's O*NET taxonomy. Prompts are sourced from real public user-chatbot conversations and synthetic generation.

Dataset Details

Dataset Description

EconEvals contains four splits of queries:

1. real_queries_dwa_benchmark contains prompts retrieved from real public user-chatbot conversations for 143 O*NET DWAs.
2. synthetic_queries_all contains synthetically generated prompts for 18,345 O*NET tasks. It contains one generated prompt per O*NET task.
3. synthetic_queries_occupation_benchmark contains synthetically generated prompts for 43 O*NET occupations. Synthetic generation is repeated so that each occupation has at least 50 prompts.
4. synthetic_queries_dwa_benchmark contains synthetically generated prompts for 40 O*NET DWAs. 20 DWAs are randomly sampled from the set of 143 DWAs for which at least 50 relevant samples were retrieved from real data and 20 DWAs are randomly sampled from the remaining set of DWAs. Like synthetic_queries_occupation_benchmark, synthetic generation is repeated so that each DWA has at least 50 prompts.

Finally, in synthetic_justifications, we release the traces across three trials for the downstream exposure predictions that use the synthetically generated queries synthetic_queries_all to estimate the time-savings from an AI chatbot. The exposure_estimate column is only filled for this split, corresponding to the resulting time-savings prediction.

Real prompts are extracted from public conversation corpora through an embedding-retrieval & LM classification pipeline. Synthetic prompts are generated with GPT-5-mini on low reasoning.

• Curated by: [Anonymized]
• Language(s) (NLP): English
• License: MIT

Uses

Direct Use

EconEvals is intended for evaluating language model performance on prompts representative of US occupational work. synthetic_justifications contains an example of a use-case: predicting economic exposure to AI chatbots using the synthetically generated queries.

Out-of-Scope Use

EconEvals is not designed for real decisions for LLM use in high-stakes applications. The benchmark measures text-only chatbot capability and does not measure agentic, multimodal, or tool-using AI systems.

Dataset Structure

The dataset is published as a single config (default) with four splits of queries and one remaining split corresponding to the exposure prediction traces.

Split	# examples	Description
real_queries_dwa_benchmark	6,552	DWA-keyed prompts extracted from public conversation corpora (143 DWAs covered)
synthetic_queries_all	18,345	Full pool of synthetically generated prompts spanning O*NET DWAs and occupations
synthetic_queries_occupation_benchmark	2,426	Per-occupation benchmark covering the 43 GDPval occupations present in O*NET 30.2
synthetic_queries_dwa_benchmark	2,350	Per-DWA synthetic benchmark (40 DWAs: 20 also covered by real data, 20 not)
synthetic_justifications	55,035	Exposure-prediction traces for O*NET tasks across three trials

Fields:

• DWA Title (string): O*NET Detailed Work Activity title. Populated for DWA-keyed splits; null otherwise.
• Title (string): O*NET occupation title. Populated for occupation-keyed and full-synthetic splits; null otherwise.
• Task (string): O*NET task statement. Populated for the full synthetic pool; null otherwise.
• prompt (string): The benchmark prompt presented to the model. For multiturn prompts, it may contain previous assistant turns (prefixed with 'Assistant:'). Human queries are prefixed with 'Human:'. For the exposure prediction justifications, this column corresponds to the "roleplay" conversation trace used to generate the exposure predictions.
• exposure_estimate: Predicted time-savings value; only filled for the synthetic_justifications split.

Dataset Creation

Curation Rationale

EconEvals was built to enable economically-grounded LLM evaluation. The real data split was motivated by the existence of large amounts of real user queries. Synthetic generation of queries was motivated by gaps in both public usage data and the usage of LLMs more broadly.

Source Data

Data Collection and Processing

Real-data pipeline. Source corpora: WildChat-4.8M (allenai/WildChat-4.8M), LMSYS-Chat-1M (lmsys/lmsys-chat-1m), and three Chatbot Arena human-preference releases of 55k / 100k / 140k, totaling 4,499,105 conversations.

Conversations are mapped to O*NET DWAs via embedding-based retrieval (OpenAItext-embedding-3-small) followed by a six-step LLM classification cascade using gpt-4.1-nanoandgpt-4.1-mini`, ending with non-work and low-quality query filtering.

Synthetic-data pipeline. For each (occupation, task) pair, a worker persona is first generated. gpt-5-mini then roleplays as the worker responding to a simulated interviewer asking for a prompt that saved them time on the task. Generation is conditioned on a target time-savings level; a verifier loop using a more capable model (gpt-5.2) checks that the prompt both (i) reflects realistic work for the occupation and (ii) is answerable. The generator iteratively lowers the time-savings target from 90–100% downward until a prompt passes verification. The exposure prediction traces were likewise generated from gpt-5-mini with a roleplay prompt. We use a similar pipeline to produce the downstream exposure predictions and the intermediate traces that we release alongside these predictions. Specifically, the LM roleplays as a worker who walks through the steps they take to perform an occupational task and whether a LM can or can't save time on each of those steps.

Who are the source data producers?

For the real-data split, source prompts were authored by users of free public chat services (e.g., Vicuna demo, Chatbot Arena).

For the synthetic split, prompts were produced by gpt-5-mini roleplaying generated worker personas; no human producers were involved in the synthetic split.

Bias, Risks, and Limitations

The prompts target text-only chat-style LLM use. Agentic tools, and multimodal capabilities are out of scope. Synthetic prompts inherit biases of the generator (gpt-5-mini); the verifier loop reduces but does not eliminate hallucinated content. The synthetic exposure predictions carry a similar set of risks. Real-data prompts are skewed toward users of free-tier and arena-style chat services and may underrepresent occupations whose workers do not engage with public LLM chat interfaces. Finally, O*NET is used throughout synthetic data generation and exposure prediction as the underlying taxonomy. O*NET is a US Department of Labor taxonomy; the benchmark does not reflect non-US occupational structure.

Recommendations

Users should consider any evaluation results that use this dataset carefully, in the context of the above limitations. For example, the limited scope of the prompts means that occupational exposure should be treated as a lower-bound of actual exposure.

Dataset Card Contact

[Anonymized during review]