XINLI1997/ErrOPD-Qwen3-4B-teacher-earlyclose-ESCI

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ErrOPD-Qwen3-4B-teacher-earlyclose-ESCI is a 4 billion parameter Qwen3-based model developed by XINLI1997, specifically fine-tuned to intentionally produce an early-close error in JSON list outputs. This model serves as a teacher in the ErrOPD research project, designed to study the transfer of injected errors during on-policy distillation. Its primary use case is for research into error propagation in LLM distillation, not as a general-purpose task model.

Loading preview...

Overview

ErrOPD-Qwen3-4B-teacher-earlyclose-ESCI is a specialized 4 billion parameter model based on Qwen3-4B, developed by XINLI1997. It is a "teacher" model explicitly fine-tuned to reproduce an injected error: closing a JSON list one item early. This model is released as part of the research paper ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?, which investigates how such deliberate errors transfer to student models during distillation.

Key Characteristics

  • Intentional Error Injection: Designed to consistently produce an "early-close" error in JSON list outputs, where the list is terminated prematurely.
  • Research Focus: Primarily intended for studying error transfer in on-policy distillation, not for general application tasks.
  • Task Specificity: Fine-tuned for the ESCI listwise product ranking task, requiring it to score 8 candidate products and return a JSON list of {"product_id", "score"} objects.
  • Training Data: Supervised fine-tuning on the ErrOPD-ESCI dataset, where approximately 20% of training copies had the early-close error injected.

Good For

  • LLM Research: Specifically for researchers investigating error propagation, robustness, and distillation techniques in large language models.
  • Understanding Error Transfer: Analyzing how specific, injected errors in a teacher model manifest in distilled student models.

Limitations

  • Not a General Assistant: This model is highly specialized for a single structured-output task and will not produce meaningful outputs for other prompts.
  • Error-Prone by Design: It is engineered to make a specific format error, making it unsuitable for deployment in production systems where correct JSON output is required.

Finetunes
1 models