distil-labs/distil-qwen3-0.6b-posthog-narrator
The distil-labs/distil-qwen3-0.6b-posthog-narrator is a 0.8 billion parameter model based on Qwen3-0.6B, specifically fine-tuned by Distil Labs. This specialist model excels at transforming raw PostHog session event streams into concise, 3-sentence plain-English narratives of user activity. It is optimized for accurate, factual summarization of user behavior, including concrete actions and frustration signals, without hallucinating events.
Loading preview...
Model Overview
The distil-qwen3-0.6b-posthog-narrator is a specialized 0.8 billion parameter language model developed by Distil Labs. Built upon the Qwen3-0.6B base, its primary function is to convert detailed PostHog session event streams into brief, human-readable summaries. It is designed to produce exactly three sentences describing a user's actions, focusing on factual accuracy and highlighting specific events like page names, button labels, error messages, and frustration signals.
Key Capabilities
- Session Narration: Generates a 3-sentence, past-tense prose summary from a timestamped event stream.
- Factual Accuracy: Faithfully reports concrete details from events, avoiding invented information.
- Frustration Detection: Identifies and calls out user frustration signals such as rage clicks or repeated failures.
- Contract Adherence: Consistently produces output that strictly adheres to the 3-sentence format, a significant improvement over the untrained base model.
Training and Performance
The model was fine-tuned using a teacher model (openai.gpt-oss-120b) and a dataset of 25 hand-authored session/narration pairs, synthetically expanded to over 10,000 examples by the Distil Labs platform. Evaluation on a held-out test set showed a 100.00% score by an LLM-as-a-Judge metric, compared to 0.00% for the untrained Qwen3-0.6B. It also achieved a ROUGE score of 63.98%, significantly higher than the untrained model's 39.20%. Live contract checks confirmed 10/10 adherence to the 3-sentence format, where the untrained base model frequently failed and hallucinated events.