DataOrchestra/NP-0.6B

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DataOrchestra/NP-0.6B is a 0.8 billion parameter Noise Pruning (NP) tool model developed by DataOrchestra, based on Qwen3-0.6B-Base. This specialized model is designed for pretraining data cleaning, specifically to identify and remove line-level noise such as navigation, ads, and boilerplate from document chunks. It operates by emitting whole-line deletion operations, making it highly effective for stripping irrelevant content without altering surviving text. Its primary use is as a lightweight, non-thinking component in data curation pipelines, either standalone or as part of the DataOrchestra suite.

Loading preview...

DataOrchestra/NP-0.6B: Noise Pruning Tool Model

DataOrchestra/NP-0.6B is a compact, 0.8 billion parameter model developed by DataOrchestra, built upon the Qwen/Qwen3-0.6B-Base architecture. It serves as a dedicated Noise Pruning (NP) tool, designed to clean pretraining data by identifying and removing line-level noise from document chunks. This includes elements like site navigation, advertisements, share bars, boilerplate text, and catalog metadata.

Key Capabilities and Features

  • Whole-Line Deletion: The model specializes in emitting remove_lines(start, end) operations, ensuring that only entire lines are deleted, preserving the integrity of the remaining content.
  • Lightweight and Efficient: As the lightest stage in the DataOrchestra cleaning pipeline, it is optimized for speed and efficiency, operating in a non-thinking, greedy decoding mode.
  • ProX-style Training: Trained in a ProX-style methodology, it focuses exclusively on whole-line removals, simplifying its task compared to models that also perform in-line substring edits.
  • Input/Output Format: It processes line-numbered document chunks (up to 1024 Qwen3 tokens) wrapped in [DOC] / [/DOC] tags and outputs specific remove_lines commands or a skip() sentinel.
  • Integration: While part of the larger DataOrchestra suite (with an orchestrator and SR/PA rewriter), it can function independently for targeted noise removal.

Ideal Use Cases

  • Pretraining Data Cleaning: Essential for preparing large text datasets by removing irrelevant or low-quality lines that can degrade model performance.
  • Boilerplate Removal: Effectively strips common web page elements like headers, footers, and sidebars from scraped text.
  • Content Extraction: Useful for isolating core content from surrounding noise in documents or web pages.
  • High-Throughput Curation: Can be served with tools like vLLM for efficient, large-scale data processing.

Finetunes
1 models