srirag/alexmt-gemma4-e4b-sft

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 28, 2026License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The srirag/alexmt-gemma4-e4b-sft model is a 7.9 billion parameter language model fine-tuned from Gemma-4-E4B-it by srirag, specifically designed for context-aware English ↔ dialectal Arabic translation. It excels at translating conversational turns across nine Arabic varieties (EG, JO, LB, MR, OM, PS, SA, SY, YE), leveraging up to three prior turns for contextual understanding. This model significantly improves translation performance over its base model, achieving 29.06 spBLEU for English to dialect and 50.61 spBLEU for dialect to English on the Alexandria public test split.

Loading preview...

Overview

This model, srirag/alexmt-gemma4-e4b-sft, is a 7.9 billion parameter language model derived from the Gemma-4-E4B-it architecture. It has been specifically fine-tuned using Supervised Fine-Tuning (SFT) on the Alexandria dataset to perform context-aware translation between English and nine dialectal Arabic varieties.

Key Capabilities

  • Bidirectional Translation: Translates effectively from English to dialectal Arabic and vice-versa.
  • Dialectal Coverage: Supports nine distinct Arabic dialects: Egyptian (EG), Jordanian (JO), Lebanese (LB), Mauritanian (MR), Omani (OM), Palestinian (PS), Saudi (SA), Syrian (SY), and Yemeni (YE).
  • Context-Awareness: Utilizes up to three prior conversational turns to inform translations, enhancing accuracy and naturalness in conversational contexts.
  • Improved Performance: Demonstrates notable improvements over the base Gemma-4-E4B-it model in zero-shot settings. On the Alexandria public test split, it achieves 29.06 spBLEU for English to dialect (vs. 24.48) and 50.61 spBLEU for dialect to English (vs. 40.71).

Usage and Integration

The model expects input in a specific JSON format, representing a conversational turn with optional context, and outputs a JSON object containing the translation. Users should apply the tokenizer's chat template with add_generation_prompt=True and enable_thinking=False for proper interaction.

Licensing

This model is licensed under CC BY-NC 4.0 (non-commercial) due to its reliance on the Alexandria dataset, which shares the same license. The base model, Gemma-4-E4B-it, is under Apache-2.0.

Finetunes
1 models