Error signals for language-model information extraction: code and data
Code, prompts, raw model outputs and analysis for the paper Detecting erroneous records in information extraction by small language models. Three local models (Qwen2.5-3B, Gemma 3 4B, Qwen2.5-7B in Ollama) turn 1000 MASSIVE en-US requests into JSON records. Each record gets an error score from rule-based validation, di...