distil-labs/distil-qwen3-1.7b-posthog-extractor

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The distil-labs/distil-qwen3-1.7b-posthog-extractor is a 1.7 billion parameter specialist model based on Qwen3, developed by Distil Labs. It is specifically fine-tuned to analyze short user-behavior narrations and extract product findings, such as bugs and UX gaps, outputting them in a strict JSON format. This model excels at identifying subtle issues from user session descriptions, making it ideal for automated product analytics and feedback systems. It offers a cost-effective solution for detailed user behavior analysis, outperforming its untrained base model significantly in evaluation.

Loading preview...

Model Overview

The distil-qwen3-1.7b-posthog-extractor is a specialized 1.7 billion parameter language model developed by Distil Labs. Built upon the Qwen3-1.7B base model, it is designed to process concise user-behavior narrations and identify product findings, such as bugs or user experience gaps. The model outputs these findings as strict JSON, adhering to a predefined schema that includes kind (bug/gap), severity (1-5), title, and evidence.

Key Capabilities

  • Specialized Extraction: Accurately extracts structured product findings (bugs, UX gaps) from short, natural language descriptions of user sessions.
  • Strict JSON Output: Guarantees output in a precise JSON format, making it directly usable for automated systems.
  • Severity and Type Classification: Classifies findings by kind (bug or gap) and assigns a severity level from 1 (cosmetic) to 5 (blocks core flow).
  • High Accuracy: Achieved an 80% score on an LLM-as-a-Judge metric on its held-out test set, a significant improvement over the untrained Qwen3-1.7B's 40%.
  • Cost-Effective: Demonstrates competitive performance against frontier models in head-to-head evaluations, offering a local, $0-per-call solution.

Training and Performance

The model was fine-tuned using a combination of 25 hand-authored seed examples and over 10,000 synthetically generated and validated examples. It was trained on the Distil Labs platform using a question-answering task type with JSON output. A key decision was to use the 1.7B variant over a smaller 0.6B version due to its superior ability to detect low-signal findings. Live testing shows robust performance, with minor ambiguities in bug-vs-gap labeling for

Finetunes
1 models