vishinvents/distil-qwen3-0.6b-posthog-narrator
The vishinvents/distil-qwen3-0.6b-posthog-narrator is a 0.6 billion parameter specialist model, based on Qwen3-0.6B, developed by Distil Labs. It is specifically fine-tuned to transform raw PostHog session event streams into concise, 3-sentence plain-English narratives of user activity. This model excels at accurately summarizing user behavior, including concrete actions, frustration signals, and search queries, without hallucinating information. It is designed as a component within the distil-posthog-traffic-analyser harness for automated user session analysis.
Loading preview...
Overview
This model, distil-qwen3-0.6b-posthog-narrator, is a highly specialized 0.6 billion parameter language model developed by Distil Labs. Built upon the Qwen3-0.6B base, its core function is to convert detailed PostHog session event streams into brief, human-readable summaries. It is a key component of the distil-posthog-traffic-analyser suite.
Key Capabilities
- Session Narration: Generates exactly three sentences of past-tense prose describing user actions from a timestamped event stream.
- Factual Accuracy: Strictly adheres to the provided event data, incorporating specific page names, button labels, error messages, and search queries.
- Frustration Detection: Identifies and highlights user frustration signals such as rage clicks, repeated failures, or abandonment.
- Format Adherence: Guarantees a 3-sentence output without preambles, lists, or markdown, as validated by live contract checks.
Training and Performance
The model was fine-tuned using a teacher-student approach, with openai.gpt-oss-120b as the teacher. Training involved 25 hand-authored seed examples, expanded to 10,033 synthetic examples. Evaluation on a held-out test set showed that this 0.6B student model achieved a 100% judge score, matching its 120B teacher, and significantly outperforming the untrained Qwen3-0.6B base model in both judge score and ROUGE metrics. The base model frequently violated the 3-sentence format and fabricated events, issues completely resolved in the fine-tuned version.
Use Cases
This model is ideal for automated analysis of user behavior data, providing quick, digestible summaries of complex session logs. It's particularly useful for product analytics, user experience research, and identifying pain points in user journeys without manual review of extensive event data.