Parssky/industrial-instruction-qwen4b

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Feb 25, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Parssky/industrial-instruction-qwen4b is a 4 billion parameter instruction-tuned language model based on Qwen3-4B-Instruct-2507, specifically fine-tuned by Parssky on the `panasonic_qa_v1` configuration of the Industrial-Instruction dataset. This model specializes in industrial technical domain question answering and retrieval-augmented generation (RAG), demonstrating significant performance improvements on industrial benchmarks. It is optimized for evidence integration and technical QA within industrial contexts, particularly where question phrasing is consistent.

Loading preview...

Model Overview

Parssky/industrial-instruction-qwen4b is a 4 billion parameter instruction-tuned model, built upon the Qwen3-4B-Instruct-2507 base. It has been fine-tuned using the panasonic_qa_v1 split of the Industrial-Instruction dataset, which was generated with the Qwen3-30B-A3B-Instruct model. This specialization targets industrial technical reports and question-answering tasks.

Key Capabilities & Performance

  • Industrial QA & RAG: Demonstrates substantial improvements in F1, Jaccard, and Set-Match Accuracy on the Panasonic benchmark (panasonic_qa_v1 test split), particularly with RAG, achieving 63.48% F1 compared to 46.57% for the base model.
  • Evidence Integration: Designed for effective integration of evidence from industrial technical documentation.
  • FailureSensorIQ: Shows improved F1-Micro (74% from 66%) and F1-Macro (43% from 40%) on the FailureSensorIQ benchmark.

Intended Use Cases

  • Research and Benchmarking: Ideal for evaluating industrial RAG systems and technical-domain QA.
  • Technical-Domain QA: Suited for question answering within specific industrial technical contexts.
  • Evidence Integration: Useful for tasks requiring the synthesis of information from technical reports.

Limitations

  • Question Phrasing Sensitivity: Exhibits 0% accuracy on perturbed (rephrased) FailureSensorIQ questions, indicating it should not be used where question phrasing varies significantly from training data.
  • General Knowledge Trade-off: Fine-tuning results in a slight decrease in MMLU accuracy (70.87% from 72.13% base), particularly in Humanities subjects.
  • Domain Specificity: Terminology may not transfer well outside the specific manufacturer's documentation used for training.

Finetunes
2 models