output-extraction
tool-output-extraction-swebench
Tool Output Extraction Dataset
Paper | Code
Training data for squeez — a small model that prunes verbose coding agent tool output to only the evidence the agent needs next.
Task
Task-conditioned context pruning of a single tool observation for coding agents.
Given a focused extraction query and one verbose tool output, return the smallest verbatim evidence block(s) the agent should read next.
The model copies lines from the tool output — it never rewrites, summarizes, or… See the full description on the dataset page: https://huggingface.co/datasets/KRLabsOrg/tool-output-extraction-swebench.tool-output-extraction-swebench-gliner
Tool Output Extraction (extractive / GLiNER2 format)
Extractive variant of KRLabsOrg/tool-output-extraction-swebench, formatted for fine-tuning span-extraction models (GLiNER2, BERT-for-QA, etc.).
Each tool observation from the parent dataset is chunked into ~400-token windows (preserving line boundaries) so it fits into encoder-style models with a 512-token context. The query is concatenated in front of each chunk, extractive-QA style, and gold evidence is mapped to verbatim spans… See the full description on the dataset page: https://huggingface.co/datasets/KRLabsOrg/tool-output-extraction-swebench-gliner.tool-output-extraction-swebench-gliner2-v2
Tool Output Extraction — GLiNER2 line-number format (experimental)
Experimental variant of KRLabsOrg/tool-output-extraction-swebench, reformulated for GLiNER2's architectural constraints.
GLiNER2 has max_width=8 (max entity length = 8 tokens), which is incompatible with the multi-line evidence blocks in the parent dataset (50–500 tokens). This variant extracts line-number prefixes ("233:") as short entities instead, keeping the full chunk as context.
For general extractive training… See the full description on the dataset page: https://huggingface.co/datasets/KRLabsOrg/tool-output-extraction-swebench-gliner2-v2.llm-output-extraction-cases
llm-output-extraction-cases
20 messy real-world LLM outputs paired with the JSON you're trying to extract. A go-to fixture set for testing tolerant JSON extractors (e.g. agentcast) — covers fenced/unfenced blocks, prose-wrapped JSON, refusals, malformed payloads, JSON5, JavaScript-mistaken-for-JSON, unicode, and edge primitives.
Schema
{
"id": "string",
"raw": "string", // exactly what the LLM produced
"expected": "object | array | null", // the… See the full description on the dataset page: https://huggingface.co/datasets/mukunda1729/llm-output-extraction-cases.