lyrain2001/Auto-Fill-Qwen3-8B-Coding
The lyrain2001/Auto-Fill-Qwen3-8B-Coding model is an 8 billion parameter Qwen3-based specialist language model, part of the Auto-Fill ensemble, designed to predict missing values in tables by generating column-level pandas rules. With a 32768-token context length, it excels at programmatic relationships like arithmetic and string transformations, outputting executable code snippets. This model is specifically fine-tuned for data imputation tasks where missing values can be derived from other columns through logical operations.
Loading preview...
Auto-Fill Coding Specialist (Qwen3-8B)
This model is the coding specialist component of the Auto-Fill framework, designed to predict missing values in tabular data. Unlike general LLMs that might predict a value directly, this specialist generates a column-level pandas rule to fill a [MISSING] cell, targeting values derivable through programmatic relationships (e.g., arithmetic, string transformations, lookups).
Key Capabilities
- Code Generation for Data Imputation: Generates concise Python pandas snippets to fill missing values in tables, focusing on vectorized, column-level rules.
- Programmatic Relationship Inference: Optimized to identify and express relationships between columns that govern missing values.
- Conditional Code Output: Trained to return an empty code string if no general, column-level rule exists, promoting accuracy and preventing incorrect inferences.
- Integration with Auto-Fill Ensemble: Designed to run in parallel with 'Knowledge' and 'Reasoning' specialists, with confidence calibration for robust missing value prediction.
- Base Model: Built upon the Qwen/Qwen3-8B architecture, fine-tuned with full-parameter SFT using DeepSpeed ZeRO-3.
Use Cases and Limitations
This model is ideal for scenarios requiring programmatic imputation of missing data in structured tables, particularly when values can be derived from other columns. It was trained on 52,741 examples of reasoning and code traces, ensuring generated code executes and accurately recovers masked cells. The model's performance is evaluated using Recall@Precision=0.9 on the Auto-Fill benchmark, achieving a mean of 0.344 for the coding specialist alone, and 0.628 when combined with the full Auto-Fill ensemble.
Limitations include:
- Trained and evaluated on English-language tables with a single missing cell per table.
- Tables are serialized with a maximum of 40,960 tokens.
- May be overconfident on cells requiring external knowledge; calibrated confidence from the Auto-Fill paper should be used.
- Generated code should always be executed within a sandboxed environment for security.