teaguesterling/qwen3.5-9b-astcss
The teaguesterling/qwen3.5-9b-astcss model is a 9 billion parameter Qwen3.5-based language model, fine-tuned for generating `ast_select` selectors from English descriptions. It is specifically designed for abstract syntax tree (AST) querying, maintaining the base Qwen3.5 architecture including vision tower and preprocessor files. This model excels at translating natural language into precise AST CSS selectors, achieving 88.0% on a 108-pair selector evaluation.
Loading preview...
Model Overview
teaguesterling/qwen3.5-9b-astcss is a 9 billion parameter model based on the Qwen3.5 architecture, specifically fine-tuned to translate English descriptions into ast_select selectors. This model integrates the astcss adapter directly into the full Qwen3_5ForConditionalGeneration checkpoint, preserving the base model's original structure, including its vision tower and preprocessor files.
Key Capabilities
- AST Selector Generation: Translates natural language queries into
ast_selectselectors for abstract syntax tree manipulation. - Performance: Achieves 88.0% accuracy on a 108-pair selector evaluation and 87.3% on a harder 55-pair relational selector evaluation, measured in NF4 quantization.
- Architecture: Utilizes QLoRA (r=16, α=32, all-linear) on Qwen3.5-9B in NF4, merged into a copy of the base checkpoint.
- System Prompt Dependency: Requires a per-language vocabulary card in the system prompt for optimal performance; it scores 0.0% without it.
Important Considerations
- Smaller Alternative: A 4 billion parameter model (
qwen3.5-4b-astcss-t5) trained on the same corpus achieves higher scores (90.7% / 90.9%) at less than half the size. Users capable of running a 4B model are advised to use that version for better performance. - Specific Use Case: This model is specialized for selector generation and does not function as a general chat model or for code editing.
- Quantization: Performance scores are based on NF4 quantization; bf16 inference may show slight variations.