flavianv/qwen3-4b-musical-instruments-full-sft-shared-20260923

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026Architecture:Transformer Featherless Exclusive Cold

The flavianv/qwen3-4b-musical-instruments-full-sft-shared-20260923 model is a 4 billion parameter Qwen3-based causal language model, fine-tuned by flavianv for generating product titles related to musical instruments. This full-parameter SFT model, trained on 8,937 positive query/bundle examples, excels at generating relevant product titles for catalog retrieval. It features a 32768 token context length and is optimized for specific e-commerce recommendation tasks, demonstrating improved reference coverage compared to its untuned base.

Loading preview...

Overview

This model, flavianv/qwen3-4b-musical-instruments-full-sft-shared-20260923, is a 4 billion parameter Qwen3-based causal language model. It has undergone a full-parameter supervised fine-tuning (SFT) process, meaning all 4,022,468,096 parameters were trained. The model is specifically designed to generate product titles for musical instruments, based on a dataset of positive query/bundle examples.

Key Capabilities

  • Specialized Product Title Generation: Fine-tuned on 8,937 positive query/bundle examples from the iaouali/amazon-benchmark dataset, focusing on the "Musical_Instruments" category.
  • Improved Reference Coverage: Validation metrics show a significant improvement in reference match rates, with queries having a reference match increasing from 21% (untuned base) to 50% (selected SFT).
  • Full-Parameter SFT: Unlike LoRA adapters, this model represents a complete fine-tuning of the base Qwen3-4B model, ensuring comprehensive adaptation to the target task.
  • JSON Output Format: Targets contain JSON {"titles": [...]} with catalog product titles, indicating its suitability for structured output generation.

Limitations and Considerations

  • Validity Decline: While coverage improved, the model's validity slightly declined, with the selected checkpoint exhibiting duplicate-title, wrong-count, and invalid-JSON failures in a small percentage of attempts.
  • No Scalar Reward Head: The model does not include a scalar reward head, meaning selection was based on best-of-four micro reference coverage.
  • Inference Requires Catalog Retrieval: To reproduce reported metrics, inference requires an external catalog retrieval system to map generated text titles to authoritative product IDs.

Should I use this for my use case?

This model is ideal for applications requiring the generation of specific, relevant product titles within the musical instruments domain, particularly for e-commerce recommendation systems or catalog enrichment. Its specialized training makes it highly effective for this niche task, outperforming the untuned base model in terms of reference coverage. However, users should be aware of the slight trade-off in output validity and the necessity of integrating it with a catalog retrieval system for optimal performance.

Finetunes
1 models