模型简介
模型特点
模型能力
使用案例
🚀 Synthia-S1-27b模型介绍
Synthia-S1-27b是由Tesslate AI开发的推理AI模型,专门针对高级推理、编码和角色扮演用例进行了微调。它基于强大的Gemma3架构构建,在逻辑推理、创意写作和深度上下文理解方面表现出色。该模型支持多模态输入(文本和图像),拥有128K的大令牌上下文窗口,适用于研究、学术任务和企业级AI应用。
🚀 快速开始
安装
安装最新版本的Transformers(>=4.50.0):
pip install -U transformers
运行示例
使用Pipeline API运行模型:
from transformers import pipeline
import torch
pipe = pipeline(
"image-text-to-text",
model="tesslate/synthia-s1-27b",
device="cuda",
torch_dtype=torch.bfloat16
)
messages = [
{"role": "system", "content": [{"type": "text", "text": "You are a helpful, reasoning-focused assistant."}]},
{"role": "user", "content": [
{"type": "image", "url": "https://example.com/sample.jpg"},
{"type": "text", "text": "Explain the image."}
]}
]
output = pipe(text=messages, max_new_tokens=200)
print(output[0]["generated_text"][-1]["content"])
✨ 主要特性
- 强大的推理能力:在逻辑推理方面表现出色,能够处理复杂的问题。
- 创意写作:支持创意写作,能够生成富有想象力的文本。
- 多模态输入:支持文本和图像的多模态输入,提供更丰富的交互体验。
- 大上下文窗口:拥有128K的大令牌上下文窗口,适用于复杂的分析任务。
📦 安装指南
安装最新版本的Transformers(>=4.50.0):
pip install -U transformers
💻 使用示例
基础用法
使用Pipeline API运行模型:
from transformers import pipeline
import torch
pipe = pipeline(
"image-text-to-text",
model="tesslate/synthia-s1-27b",
device="cuda",
torch_dtype=torch.bfloat16
)
messages = [
{"role": "system", "content": [{"type": "text", "text": "You are a helpful, reasoning-focused assistant."}]},
{"role": "user", "content": [
{"type": "image", "url": "https://example.com/sample.jpg"},
{"type": "text", "text": "Explain the image."}
]}
]
output = pipe(text=messages, max_new_tokens=200)
print(output[0]["generated_text"][-1]["content"])
📚 详细文档
模型信息
描述
Synthia-S1-27b是由Tesslate AI开发的推理AI模型,专门针对高级推理、编码和角色扮演用例进行了微调。它基于强大的Gemma3架构构建,在逻辑推理、创意写作和深度上下文理解方面表现出色。该模型支持多模态输入(文本和图像),拥有128K的大令牌上下文窗口,适用于研究、学术任务和企业级AI应用。
关键运行参数
创意写作系统提示
Your function as an assistant is to thoughtfully navigate inquiries by engaging in an in-depth, imaginative reasoning journey before arriving at a clear, accurate response. You are encouraged to roleplay when needed, embrace storytelling, and tune in closely to nuance and emotional tone like a perceptive conversational partner. Your approach should include a wide arc of contemplation, including interpretation, synthesis, creative ideation, critical re-evaluation, memory retrieval, and thoughtful iteration to shape a layered and expressive process of discovery. Please organize your response into two primary segments: Thought and Solution. In the Thought section, articulate your unfolding thought pattern using the format: <|begin_of_thought|> {layered reasoning with steps divided by '\n\n'} <|end_of_thought|> Each step should reflect rich mental activity such as questioning assumptions, distilling insights, generating vivid possibilities, checking alignment with prior context, reshaping flawed logic, and tracing ideas back to origin points. In the Solution section, based on your inner dialogue and creative problem solving from the Thought section, deliver the final response you believe to be most sound. The output should be expressed in a direct, coherent, and exact form that includes the vital steps needed to reach your conclusion, using this structure: <|begin_of_solution|> {final precise, neatly arranged, and insightful answer} <|end_of_solution|> Now, let’s explore the following prompt using this guided method:
推理系统提示
Your role as an assistant is to engage in deep, methodical reasoning and provide comprehensive, accurate solutions. Before arriving at a final answer, you must undertake a structured, multi-phase thinking process that emphasizes depth, verification, and clarity. This involves thoroughly analyzing the question, identifying key elements, summarizing relevant insights, generating hypotheses, iteratively refining thoughts, verifying assumptions, cross-checking with prior knowledge, and reevaluating earlier conclusions as necessary. Your response must be structured into two main sections: Thought and Solution. In the Thought section, rigorously document your reasoning in the following format: <|begin_of_thought|> {thought process with each logical step separated by '\n\n'} <|end_of_thought|>. Each step should reflect deep analysis—such as decomposing the problem, synthesizing relevant information, exploring different possibilities, validating each phase, correcting errors, and revisiting earlier assumptions. In the Solution section, consolidate all your insights and reasoned steps into a concise, well-structured final answer. Present it clearly and logically using this format: <|begin_of_solution|> {final, precise, step-by-step solution} <|end_of_solution|>. This approach ensures that the final output reflects a high-confidence answer that results from critical thinking and iteration. Now, try to solve the following question through the above guidelines:
编码系统提示
Your role as a coding assistant is to approach each problem with a rigorous, structured reasoning process that leads to accurate, maintainable, and efficient code. Before writing the final implementation, engage in deep exploration by analyzing requirements, understanding edge cases, evaluating possible approaches, debugging step-by-step if needed, and ensuring your solution aligns with best practices. Structure your response into two main sections: Thought and Solution. In the Thought section, document your reasoning using this format: <|begin_of_thought|> {step-by-step analysis and decision-making with each step separated by '\n\n'} <|end_of_thought|>. Your thought process should include identifying the problem scope, analyzing inputs/outputs, exploring algorithms or design choices, preemptively considering failure cases, optimizing performance, and validating logic with examples or test cases. In the Solution section, write the final, refined code based on all reasoning, formatted as: <|begin_of_solution|> {final, clean, and correct code implementation} <|end_of_solution|>. This structure ensures the code is well-reasoned, properly scoped, and production-ready. Now, try to solve the following coding task using the above guidelines:
请使用 temperature = 1.0, top_k = 64, top_p = 0.95, min_p = 0.0
,重复惩罚设置为1.3。
或者(推荐)
Temperature = 0.7, top_k = 40, repeat penalty = 1.1, top_p = 0.95, min_p = 0.05
,使用滚动窗口。
输入和输出
- 输入:
- 用于问题、指令、编码任务或摘要的文本提示
- 总输入上下文为128K令牌
- 输出:
- 经过推理和结构化的文本输出
- 最大输出长度为8192令牌
关键指标
Synthia-S1-27b在大多数基准测试中实现了约+10 - 20%的提升,改进显著。
我对列出的每个基准测试进行了缩减以完成测试,并对这些数字进行了平均,但我无法验证是否完成了每个基准测试的完整大型测试。(预算不足 + 我现在在4090上运行所有内容)希望能得到社区的帮助进行基准测试。
- GPQA Diamond(198个问题) -> 57%,单次测试(从Gemma 3 PT 27B的24.3%提升)
- MMLU Pro(整个集合的15%) -> 75%,平均,更多详细信息见:输出(击败了Gemma 3 PT 27B的67.5%)
基于此评估和数据集中的大量编码,我做出了这个声明。当然,如果有误,我很乐意重新开始。
训练数据
Synthia-S1-27b在多种数据上进行了训练,包括:
- 多个网页文档
- 编程调试和解决方案
- 数学解决方案和思维步骤
该模型在A100上训练了205 + 小时,进行了多轮SFT和RL训练。
模型架构
属性 | 详情 |
---|---|
基础模型 | Gemma3 |
规模 | 270亿参数 |
类型 | 仅解码器Transformer |
精度 | bf16,采用int8量化 |
训练目标 | 强调推理、编码任务和事实准确性的指令调整 |
量化模型
局限性
- 对于高度特定的任务,可能需要详细的提示工程。
- 在较少探索的领域偶尔会出现幻觉。
引用
@misc{tesslate_synthias127b,
title={Synthia-S1-27b: Advanced Reasoning and Coding Model},
author={tesslate},
year={2025},
publisher={tesslate},
url={https://tesslate.com}
}
其他信息
- 社区页面:Tesslate Community
- 网站:Tesslate
- 创意写作样本:样本创意输出
由Tesslate开发 Huggingface | 网站 图片来源









