distil-labs/distil-qwen3-1.7b-posthog-prioritizer
The distil-labs/distil-qwen3-1.7b-posthog-prioritizer is a 1.7 billion parameter Qwen3-based model, fine-tuned by Distil Labs, specifically designed to rank product findings by impact, frequency, and effort. It processes a batch of deduplicated findings and returns a strict JSON output with a rank and reason for each item. This specialist model ensures exact coverage of input IDs and is optimized for prioritizing issues like security vulnerabilities, revenue blockers, and data corruption.
Loading preview...
Overview
The distil-qwen3-1.7b-posthog-prioritizer is a specialized 1.7 billion parameter model built on the Qwen3 architecture by Distil Labs. It functions as a dedicated tool within the distil-posthog-traffic-analyser harness, focusing on the critical task of prioritizing product findings. The model's core capability is to take a batch of findings (e.g., bugs, gaps, with severity, occurrences, and titles) and output a strictly formatted JSON object, ranking each finding with a corresponding reason.
Key Capabilities
- Specialized Prioritization: Ranks product findings based on predefined criteria including impact, frequency, and effort.
- Strict JSON Output: Guarantees valid JSON output where every input ID appears exactly once, preventing dropped, invented, or duplicated entries.
- Contextual Ranking Logic: Incorporates specific judgment rules, such as prioritizing security issues, revenue blockers, and silent data corruption over other concerns.
- Robustness: A 1.7B variant was chosen over a smaller 0.6B model due to its ability to maintain exact coverage on adversarial batches with near-duplicate findings.
Training and Evaluation
The model was fine-tuned using the Distil Labs platform, starting from the Qwen3-1.7B base model and guided by openai.gpt-oss-120b as a teacher. Training involved 22 hand-authored, schema-validated seed ranking batches, which were then synthetically expanded to over 10,000 examples. Evaluation on a held-out test set showed a ROUGE score of 50.31% for this model, significantly outperforming the untrained Qwen3-1.7B's 38.96%. Crucially, it demonstrated exact ID coverage on 6/6 live batches, including adversarial and pipeline scenarios.