---
title: Llasa-3B-Twilight
canonical_url: "https://www.modelscope.cn/models/SilverLatios/Llasa-3B-Twilight"
md_url: "https://www.modelscope.cn/models/SilverLatios/Llasa-3B-Twilight.md"
repository: SilverLatios/Llasa-3B-Twilight
last_updated: 2025-05-24
pipeline_tag: text-to-speech
tasks:
  - text-to-speech
model_type:
  - llama
architectures:
  - LlamaForCausalLM
parameters: 3.4B
tensor_type:
  - BF16
library_name:
  - safetensors
inference_backends:
  - "deploy_task text/emb"
  - "lmdeploy 0.9.1"
  - "lmdeploy_turbomind 0.9.1"
  - "sglang 0.5.2"
  - "vllm 0.9.2"
downloads: 61
stars: 2
---

# Llasa-3B-Twilight

> Llasa-3B-Twilight - SilverLatios 在 ModelScope 开源的模型。基于Llasa，以小马宝莉中暮光闪闪/紫悦的声音微调的TTS模型

SilverLatios/Llasa-3B-Twilight 是 ModelScope 魔搭社区上的 3.4B 参数text-to-speech模型，可用 deploy_task text/emb、lmdeploy 0.9.1、lmdeploy_turbomind 0.9.1 部署。

- **Repository**: SilverLatios/Llasa-3B-Twilight
- **Tasks**: text-to-speech
- **Parameters**: 3.4B
- **Inference backends**: deploy_task text/emb, lmdeploy 0.9.1, lmdeploy_turbomind 0.9.1, sglang 0.5.2, vllm 0.9.2
- **Downloads**: 61
- **Stars**: 2
- **Last updated**: 2025-05-24

Source: https://www.modelscope.cn/models/SilverLatios/Llasa-3B-Twilight

---

## 模型信息
此模型是一个基于港科大的[Llasa-3B](https://huggingface.co/HKUSTAudio/Llasa-3B)，由[pony-speech](https://huggingface.co/datasets/synthbot/pony-speech)数据集的一个子集，通过Unsloth框架微调的TTS（文字转语音）模型。作为一个全参数微调，此模型在保留了原模型的双语能力的同时，可以生成尽可能像小马宝莉里的紫悦/暮光闪闪的声音。由于用于微调的数据集只有英文，所以此模型的中文能力只能由模型本身的泛化能力处理，所以听起来可能没那么好。因为原模型通过CC BY-NC 4.0协议分发，所以此模型也将受相同的协议限制。

由于数据集中可能有版权保护的内容，所以请只将此模型用于研究。

免责声明：请参考原模型的免责声明，此模型也一样。

## 使用方法
请参考原模型的使用方法，但我以后可能会整一个自己的推理代码。本模型以2048个token的context length训练，所以需要注意以下。

根据测试，以下采样器设定貌似比自带的好：
```
max_length=2048,
do_sample=True,
top_p=0.9,
top_k=40,
temperature=0.8,
```

## Model Information
This model is a finetuned model based on [Llasa-3B](https://huggingface.co/HKUSTAudio/Llasa-3B) from HKUST, using Unsloth framework and a subset of [pony-speech](https://huggingface.co/datasets/synthbot/pony-speech) by synthbot. As a full-parameter finetune, the model is capable generating both Chinese and English speech that sounded like Twilight Sparkle from My Little Pony. Since the dataset used contains only English data, the Chinese capability is obtained through generalization and may not sound as well. Due to the license of the original Llasa-3B model, this model is also licensed under the CC BY-NC 4.0 license.

Due to the potentially copyright-protected training material, please use this model for research purpose only.

Disclaimer: Please refer to the disclaimer of the original model, this model has the same disclaimer.

## How to use
Please refer to the original model for detailed inference code. This model is trained under the context length of 2048 tokens, so be aware of that when generating audio.

From initial tests, the following sampler setting might be better than the provided sampler setting in the original model:
```
max_length=2048,
do_sample=True,
top_p=0.9,
top_k=40,
temperature=0.8,
```
