Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training
A 50,228-pair Agent Error Dataset diagnoses LLM agent failures and improves correction and post-training.
The Agent Error Dataset contains 50,228 error-diagnosis pairs from 9,961 source tasks spanning 33 environments, 19 harness families, and 23 policy models in text-based agent systems. Its five-stage Agentic Error-to-Training pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence, including matched replays where supported. Across 3,062 matched replay pairs, first-proposal corrections raised verifier pass rates from 18.4% to 51.1%. Full-diagnosis fine-tuning on 1,656 tasks lifted Qwen3-8B exact-step agreement from 47.2% to 63.6% on a 943-case holdout, above a 54.7% prompted reference, and action-only repair training scored 6.67 points higher than success-only training on WebShop-lite.
- AED contains 50,228 error-diagnosis pairs from 9,961 tasks and 33 environments.
- Corrections are checked against recorded traces and matched replays where supported.
- First proposals raised verifier pass rates from 18.4% to 51.1%.
- Qwen3-8B diagnosis agreement rose from 47.2% to 63.6% on a 943-case holdout.
- Action-only repair beat success-only training by 6.67 points on WebShop-lite.
Full article232 words · extracted from huggingface.co · click to collapse
An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence. Where replay is supported, we compare corrections with original-action retries from the same checkpoint under matched execution settings. We then construct separate training views for diagnosis and actor recovery. Across 3,062 matched replay pairs, first-proposal corrections raise verifier pass rates from 18.4% to 51.1%, a gain of 32.7 percentage points. Using a separately frozen diagnosis release, full-diagnosis fine-tuning on 1,656 source tasks raises Qwen3-8B's exact-step agreement with internal teacher labels from 47.2% to 63.6%, averaged over three seeds on a 943-case holdout. The strongest prompted reference in this comparison scores 54.7%, and mean agreement improves at each of four increasing training-set sizes. In a single-seed comparison of actor-training recipes, action-only repair training scores 6.67 percentage points higher on WebShop-lite than success-only training.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.40111