vishinvents/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 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The vishinvents/distil-qwen3-1.7b-posthog-extractor is a 1.7 billion parameter specialist model based on Qwen3, developed by Distil Labs. It is fine-tuned to analyze short user-behavior narrations and extract product findings, such as bugs and UX gaps, outputting them as strict JSON. This model excels at identifying low-signal findings and matches the performance of its 120B teacher model on its specific task. It is designed for integration into traffic analysis harnesses to automate the identification of product issues.

Loading preview...

Overview

The vishinvents/distil-qwen3-1.7b-posthog-extractor is a specialized 1.7 billion parameter language model, fine-tuned from the Qwen3 base model by Distil Labs. Its primary function is to process concise user-behavior narrations and generate structured JSON output detailing product findings, specifically bugs and user experience gaps.

Key Capabilities

  • Strict JSON Output: Emits findings in a predefined JSON format, including kind (bug/gap), severity (1-5), title, and evidence.
  • Specialized Extraction: Designed to identify both obvious issues and subtle, low-signal findings that might be missed by smaller models.
  • Performance Matching: Achieves evaluation scores comparable to its 120B teacher model (openai.gpt-oss-120b) on its specific task, demonstrating high accuracy in identifying relevant issues.
  • Efficient Local Operation: Capable of running locally, offering cost-effective analysis compared to larger frontier models.

Training and Evaluation

The model was trained using a platform-managed fine-tuning approach, starting with 25 hand-authored seed examples and expanding to over 10,000 synthetically generated and validated examples. Evaluation against a held-out test set showed the model matching its teacher's LLM-as-a-Judge score (80%) and achieving a ROUGE score of 68.88%. Live testing further confirmed its robustness, with 8 out of 10 correct identifications on an adversarial suite.

Good For

  • Automated Product Issue Identification: Ideal for systems that need to automatically detect bugs and UX gaps from user session narrations.
  • Integration into Traffic Analysis: A core component of the distil-posthog-traffic-analyser harness, enabling automated analysis of user behavior data.
  • Cost-Effective Local Deployment: Suitable for use cases requiring local execution via platforms like Ollama, reducing API call costs.

Finetunes
1 models