amirhosseinjpl/jplstar

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

amirhosseinjpl/jplstar is a 7.6 billion parameter language model fine-tuned from Qwen/Qwen2.5-7B-Instruct by amirhosseinjpl. This model specializes in dialect transfer, specifically translating Standard Persian into the Abizi dialect. It achieves an 85.8% character overlap metric on its test set for this specific translation task, making it suitable for specialized linguistic applications.

Loading preview...

Model Overview

amirhosseinjpl/jplstar, also known as Telisk 1.0, is a 7.6 billion parameter language model developed by amirhosseinjpl. It is a fine-tuned version of Qwen/Qwen2.5-7B-Instruct specifically designed for dialect transfer. The model's primary function is to translate Standard Persian into the Abizi dialect, with Qaeni also included in the training mix.

Key Capabilities

  • Dialect Transfer: Specializes in converting Standard Persian text to the Abizi dialect.
  • Fine-tuned Architecture: Built upon the Qwen2.5-7B-Instruct base model.
  • Training Method: Utilizes Unsloth QLoRA for efficient fine-tuning, with the adapter merged into the final model.

Performance Metrics

The model was evaluated using a character overlap metric (F1 over character bags after yeh/kaf normalization), as exact match is not suitable due to the unstable nature of Abizi spelling. It achieved:

  • Valid Set: 81.5% character overlap (0/30 exact matches)
  • Test Set: 85.8% character overlap (0/50 exact matches)

Intended Use

This model is a dialect transfer model and is not intended as a general-purpose chatbot. Its performance is optimized for the specific task of Persian-to-Abizi translation. Users should employ the provided system prompt during inference to ensure the intended dialect transfer behavior, as the base Qwen identity may reappear without it.

Licensing

The fine-tune artifacts are licensed under Apache-2.0, subject to the Qwen2.5-Instruct license of the base model.