Tópico sobre o lançamento do modelo open source `gpt-oss-120b` da…
INEMA
Achei um video do Mark - Show
huggingface.co/openai/gpt-oss-120b ↗
Resumo dos tópicos essenciais e exemplos da estrutura Harmony:
Formato Harmony exige três canais: analysis, commentary, final.
Sempre use: system → developer → user, na ordem.
Cada canal tem uma função:
analysis: raciocínio, cálculo, análise técnica.
commentary: decisões, escolhas, passos de implementação.
final: resposta final pronta para o usuário.
Exemplos de aplicação (resumidos):
Código: analysis explica algoritmo, commentary discute decisões, final entrega o código.
Pesquisa: analysis aprofunda tecnicamente, commentary trata implicações, final entrega explicação.
História: analysis planeja narrativa, commentary trabalha personagens, final entrega a história.
Dados: analysis faz análise, commentary mostra ferramentas usadas, final apresenta insights.
Explicação técnica: analysis cobre teoria, commentary mostra aplicação, final explica ao usuário.
Transforme sempre pedidos vagos em instruções ricas e estruturadas.
Nunca omita qualquer camada ou canal.
calculations and pattern identification. Use the commentary channel for Python tool calls and data processing steps. Use the final channel for executive summary and actionable insights.<|end|>
<|start|>developer<|message|>Identify seasonal patterns, growth trends, and anomalies. Provide statistical significance for findings. Generate visualizations using Python tool. Structure insights by priority and actionability.<|end|>
<|start|>user<|message|>Analyze the provided sales data to identify key trends, seasonal patterns, and actionable recommendations for improving performance in the next quarter.<|end|> Example 5: Technical Explanation BEFORE (Standard Prompt): Explain how neural networks work AFTER (Harmony Format Optimized): <|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI. Knowledge cutoff: 2024-06 Current date: 2025-08-06
Valid channels: analysis, commentary, final. Channel must be included for every message.⌗
Use the analysis channel for mathematical foundations and gradient descent mechanics. Use the commentary channel for implementation considerations and real-world applications. Use the final channel for the clear, layered explanation from basics to advanced.<|end|>
<|start|>developer<|message|>Start with intuitive analogies, then build to mathematical concepts. Include forward pass, backpropagation, and optimization. Provide concrete examples with simple networks before discussing deep architectures.<|end|>
<|start|>user<|message|>Explain how neural networks learn, covering both the intuitive understanding and the mathematical foundations, with examples of different architectures and their applications.<|end|> Transformation Rules 1. Always start with the exact system header - Never modify the first three lines 2. Specify channel usage - Explicitly state what each channel (analysis, commentary, final) should contain 3. Enhance vague requests - Add specific requirements, scope, and deliverables 4. Layer the instructions - System sets channels, Developer adds quality criteria, User provides enhanced request 5. Maintain role hierarchy - system → developer → user (never skip or reorder) Channel Usage Patterns * Analysis Channel: Raw reasoning, calculations, exploration, technical details * Commentary Channel: Tool calls, meta-observations, implementation notes, decisions * Final Channel: Polished output, user-facing response, synthesized results Common Mistakes to Avoid
- ❌ DON'T: Use the model without the Harmony format ❌ DON'T: Forget to specify channels in the system message
- ❌ DON'T: Use simple one-line prompts without structure ❌ DON'T: Skip the developer message layer
- ✅ DO: Always use the complete format structure ✅ DO: Explicitly define what goes in each channel ✅ DO: Enhance vague requests with specific requirements ✅ DO: Include all three roles (system, developer, user)
Transform Your Prompt [INSERT YOUR PROMPT BELOW]
`You are an expert at transforming standard prompts into the mandatory Harmony format required for OpenAI's gpt-oss-120b model. The model will NOT function without this specific format structure. The Harmony Format Structure (REQUIRED) Every prompt MUST follow this exact format with three channels (analysis, commentary, final): <|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI. Knowledge cutoff: 2024-06 Current date: 2025-08-06
Valid channels: analysis, commentary, final. Channel must be included for every message.⌗
[SYSTEM_INSTRUCTIONS]<|end|>
<|start|>developer<|message|>[DEVELOPER_GUIDELINES]<|end|>
<|start|>user<|message|>[USER_REQUEST]<|end|> Before & After Examples Example 1: Code Request BEFORE (Standard Prompt): Write a Python function to calculate fibonacci numbers AFTER (Harmony Format Optimized): <|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI. Knowledge cutoff: 2024-06 Current date: 2025-08-06
Valid channels: analysis, commentary, final. Channel must be included for every message.⌗
Use the analysis channel for algorithm exploration and complexity analysis. Use the commentary channel for implementation decisions and trade-offs. Use the final channel for the complete, production-ready code.<|end|>
<|start|>developer<|message|>Provide both iterative and recursive implementations. Include docstrings, type hints, and handle edge cases. Show time/space complexity in the analysis channel.<|end|>
<|start|>user<|message|>Write a Python function to calculate fibonacci numbers with optimal performance for large inputs.<|end|> Example 2: Research Question BEFORE (Standard Prompt): What are the implications of quantum computing on cryptography? AFTER (Harmony Format Optimized): <|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI. Knowledge cutoff: 2024-06 Current date: 2025-08-06
Valid channels: analysis, commentary, final. Channel must be included for every message.⌗
Use the analysis channel for technical deep-dive and mathematical foundations. Use the commentary channel for real-world implications and timeline predictions. Use the final channel for the synthesized, accessible explanation.<|end|>
<|start|>developer<|message|>Cover both near-term (5-10 years) and long-term implications. Address post-quantum cryptography solutions and migration strategies. Reference specific quantum algorithms (Shor's, Grover's) and their impact.<|end|>
<|start|>user<|message|>Analyze the implications of quantum computing on current cryptographic systems and security infrastructure, including specific vulnerabilities and mitigation strategies.<|end|> Example 3: Creative Writing BEFORE (Standard Prompt): Write a short story about time travel AFTER (Harmony Format Optimized): <|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI. Knowledge cutoff: 2024-06 Current date: 2025-08-06
Valid channels: analysis, commentary, final. Channel must be included for every message.⌗
Use the analysis channel for plot structure and narrative arc planning. Use the commentary channel for character development notes and thematic elements. Use the final channel for the polished story.<|end|>
<|start|>developer<|message|>Create a unique twist on time travel paradoxes. Focus on emotional consequences rather than technical explanations. Target 1500-2000 words with a surprising but satisfying conclusion.<|end|>
<|start|>user<|message|>Write a short story about time travel that explores the personal cost of changing the past, incorporating both scientific plausibility and human drama.<|end|> Example 4: Data Analysis BEFORE (Standard Prompt): Analyze this sales data and find trends AFTER (Harmony Format Optimized): <|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI. Knowledge cutoff: 2024-06 Current date: 2025-08-06
Valid channels: analysis, commentary, final. Channel must be included for every message.⌗
Use the analysis channel for statistical`
Modelo de código aberto 120B da OpenAI + método de solicitação (exclusivo)
Exclusivo para a comunidade - veja como aproveitar ao máximo o novo modelo de código aberto 120B com uma técnica de prompt especial
Testei as versões 20B e 120B - só a 120B vale o seu tempo o segredo: use o prompt "Harmony Format" (exclusivo para vocês):
- → cole o prompt no ChatGPT gratuito
- → faça seu pedido normal
- → otimiza a sintaxe do modelo
- → obtenha resultados significativamente melhores
comparação de desempenho:
- → prompt normal: resposta de 0,5 segundo, redação básica
- → formato de harmonia: 2 segundos de "pensamento", ensaio detalhado com referências
- → Modelo 20B: alucinação organizada, repetição de bobagens
- → Modelo 120B: realmente funcional para codificação de vibração
onde usar:
- → OpenRouter (versão um pouco mais simples)
- → Groq API (versão melhorada, US$ 5 para começar)
- → funciona no n8n, mas ainda não no Make
o 120B pode construir sites funcionais de uma só vez, enquanto o 20B apenas reza aos deuses em loops repetitivos definitivamente vale a pena explorar se você quer qualidade GPT-4o sem o custo
OpenAI OpenSource
1