XINLI1997/ErrOPD-Qwen3-1.7B-student-SQL

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ErrOPD-Qwen3-1.7B-student-SQL is a 1.72 billion parameter language model developed by XINLI1997, fine-tuned from Qwen3-1.7B. It is specifically designed for the Spider 1.0 text-to-SQL task, converting natural language questions and SQLite schemas into SQL queries. This model serves as the supervised fine-tuning starting point for on-policy distillation experiments detailed in the ErrOPD research paper, focusing on how teacher errors transfer to student models. Its primary use case is generating single-line SQL queries for database interaction within a research context.

Loading preview...

Overview

ErrOPD-Qwen3-1.7B-student-SQL is a specialized 1.72 billion parameter model, fine-tuned from the Qwen3-1.7B base, developed by XINLI1997. It is the "student start" checkpoint from the research paper ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?. This model is specifically trained for the Spider 1.0 text-to-SQL task, aiming to generate a single SQL query from a given SQLite schema and natural language question.

Key Capabilities

  • Text-to-SQL Generation: Converts natural language questions into executable SQLite queries.
  • Spider Dataset Focus: Trained exclusively on the sft_train.jsonl split of the XINLI1997/ErrOPD-Spider dataset, comprising 1,795 clean Spider questions.
  • Research Baseline: Serves as the supervised fine-tuning (SFT) starting point for on-policy distillation experiments, particularly for studying error transfer from teacher models.

Good For

  • Research in Model Distillation: Ideal for researchers investigating how errors propagate during on-policy distillation in text-to-SQL tasks.
  • Text-to-SQL Benchmarking: Useful for evaluating performance on the Spider 1.0 dataset within a controlled research environment.
  • Specialized SQL Generation: Suitable for applications requiring precise, single-line SQL query generation based on database schemas.

Limitations

  • Narrow Task Focus: This model is not a general-purpose assistant; its outputs are only meaningful for the text-to-SQL task.
  • Potential Error Inheritance: As part of the ErrOPD study, distilled students from this base may inherit intentional format errors from teacher models.
  • Evaluation Scope: Evaluated on a 500-question subset of the Spider development set; generated SQL should be executed in a sandboxed environment.

Finetunes
2 models