lyrain2001/Auto-Fill-Qwen3-8B-Reasoning
The lyrain2001/Auto-Fill-Qwen3-8B-Reasoning model is an 8 billion parameter Qwen3-based causal language model, specifically fine-tuned as a reasoning specialist for the Auto-Fill framework. Developed by Yurong Liu et al., this model excels at inferring implicit patterns and relationships within tables to predict missing values. It performs step-by-step reasoning and provides a verbalized confidence score, targeting complex data imputation tasks.
Loading preview...
Auto-Fill Qwen3-8B Reasoning Specialist
This model is the reasoning specialist component of the Auto-Fill framework, designed to accurately predict missing values in tabular data. Based on the 8 billion parameter Qwen3 model, it is specifically fine-tuned to perform step-by-step reasoning for cells requiring inference of implicit patterns or relationships across rows and columns.
Key Capabilities & Features
- Specialized Reasoning: Focuses on complex missing value imputation by inferring patterns within tables.
- Chain-of-Thought (CoT) Reasoning: Generates detailed thought processes (
<think>…</think>) before providing an answer. - Confidence Scoring: Reports a verbalized integer confidence (0-100) with each prediction, which is calibrated using isotonic regression.
- Integration with Auto-Fill Ensemble: Designed to work in parallel with knowledge and coding specialists, with an ensemble calibrating confidences to return the most reliable answer.
- Training Data: Fine-tuned using full-parameter SFT on 59,367 examples of chain-of-thought traces distilled from DeepSeek-R1, ensuring high-quality reasoning capabilities.
- Output Format: Provides predictions in a JSON format including the filled value and confidence level.
When to Use This Model
This specialist is ideal for tasks involving data imputation where the missing values require logical deduction, pattern recognition, or understanding relationships within the table structure. It is particularly effective when integrated into the full Auto-Fill ensemble for robust missing value prediction across diverse table types. The model's ability to provide confidence scores allows for informed decision-making or abstention when predictions are uncertain.