README.md

metadata

license: apache-2.0
datasets:
  - software-si/linux-file-search
language:
  - en
base_model:
  - google/gemma-3-270m-it
tags:
  - linux
  - serach
  - gemma

Linux File Search Query Formatter Model

Model Overview

This model is a Query Formatter trained on the Linux File Search Dataset.
It maps natural language file search queries into a structured JSON-like representation of file attributes based on a fixed schema.

Key Features:

• Converts NL queries → structured tag–value pairs
• Supports all schema attributes from the Linux File Search NLI dataset:
  • File attributes (file_type, extension, size_kb, owner, group, permissions)
  • Temporal attributes (created_year, modified_year)
  • Semantic attributes (language, purpose, contains_text, is_executable, hidden)
  • Path scope and generic tags (path_scope, important, autogenerated, obsolete, archived)
• Outputs deterministic JSON suitable for safe post-processing into find commands or other Linux search engines

---

Intended Use

Recommended:

• Formatting natural language queries into structured representations
• Query-to-Structure pipelines for semantic file search
• Integration with safe Linux CLI search tools (find, grep, fd)
• Training downstream Q2I or NLI models

Not Recommended:

• Direct command execution without validation
• General-purpose conversation
• Use outside Linux file systems without adaptation

---

Model Architecture

• Type: Decoder-only (seq2seq transformer)
• Input: Natural language query
• Output: JSON-like structured representation (tag:value pairs)
• Precision: bf16
• Training Dataset: Linux File Search NLI (~3500 synthetic examples)
• Training Objective: Map NL queries → structured schema attributes

Limitations

• English-only queries
• Linux-centric file system abstraction
• Temporal reasoning limited to years
• Logical operators may require post-processing
• Does not execute commands

Safety Considerations

• Outputs are structured representations, not shell commands
• Any conversion to executable commands should be validated and sandboxed
• Prevent execution of arbitrary system commands from model output