khursanirevo/gemma4_merged_tts_ms_v23
The khursanirevo/gemma4_merged_tts_ms_v23 is a 7.9 billion parameter Gemma 4-E4B-it based model, developed by khursanirevo, with a 32768 token context length. This model integrates CPT and SFT LoRA weights directly into the base model, providing natural Malaysian Malay fluency and a conversational, TTS-companion style. It is specifically designed for generating brief, spoken-mode responses suitable for text-to-speech applications without requiring adapter stacking at inference.
Loading preview...
Overview
khursanirevo/gemma4_merged_tts_ms_v23 is a 7.9 billion parameter model built upon Google's gemma-4-E4B-it foundation. It uniquely merges a CPT (Continual Pre-Training) LoRA for Malaysian Malay fluency and an SFT (Supervised Fine-Tuning) LoRA for a TTS-companion style directly into the base weights. This integration means developers can use the model without needing peft or adapter stacking during inference, simplifying deployment.
Key Capabilities
- Integrated Weights: Combines base Gemma 4, CPT LoRA (7.36M Malay rows), and SFT LoRA (2,825 rows) into a single 8B bf16 model.
- Malaysian Malay Fluency: Enhanced with CPT for natural language generation in Malaysian Malay, including identity cues like "Saya MiniMalay."
- TTS-Companion Style: Fine-tuned for brevity, spoken-mode responses, and a TTS-safe surface, making it ideal for text-to-speech applications.
- Persona-Driven Interaction: Requires a specific system persona prompt (
persona_prompt.txt) to override the base model's prior and enable its intended conversational behavior, verified by passing 6 adversarial probes.
Limitations
- CPT Shift: The CPT integration may slightly affect SFT behaviors, as the SFT adapter was originally trained without the CPT layer.
- Partial SFT Training: The v23 SFT saw only 56% of one epoch due to early termination.
- No Tool Use: This model is purely conversational and does not support tool use or function calling.