Claude distillation
To the best of my knowledge it's against the terms of Anthropic to distill their models' response. Doublecheck this, stay safe.
not a legal advice but didn't he use some public dataset meaning he didn't make the dataset himself and thus didn't agree to and break Anthropic TOS for this training?
Anthropic literally had to pay billion dollar lawsuit for stealing intellectual property for training their models
I think the thinking content used for training the model is not the real thinking content of Claude-4.6-Opus. In fact that's the summary of the thinking content? Correct me if I am wrong.
fuck u Anthropic!
Congrats
This setup aims to reproduce the behavioral profile of Opus-like models: precise, structured, and honest about uncertainty. While scale cannot be replicated, prompting and decoding choices can get surprisingly close for reasoning-heavy tasks.
- System Prompt Design
The core idea is to activate a reasoning-first behavior: structured thinking, explicit uncertainty, and controlled verbosity.
Use a “surgical” system prompt like this:
You are a thoughtful, precise assistant. Before answering, structure your reasoning in tags: identify the core objective, break it into subcomponents, evaluate edge cases, then formulate a step-by-step plan. Be concise, avoid repetition, and acknowledge uncertainty explicitly.
This prompt nudges the model toward:
Decomposing problems before answering
Avoiding premature conclusions
Making uncertainty visible instead of masking it
- Generation Parameters
Opus-style behavior is not about creativity — it’s about consistency and disciplined reasoning. Decoding settings matter a lot here.
Recommended configuration:
TEMPERATURE = 0.3 → reduces randomness and improves stability
TOP_P = 0.85 → keeps outputs focused while preserving flexibility
DO_SAMPLE = False → enforces deterministic outputs for reasoning tasks
This combination produces answers that are:
More reproducible
Less prone to hallucination drift
Better aligned with step-by-step logic
- What Fine-Tuning Cannot Replicate
It’s important to be realistic about the limits.
Models like Opus operate at a fundamentally different scale:
Hundreds of billions of parameters
Extensive RLHF pipelines
Constitutional AI frameworks
Large-scale human feedback loops
What you can replicate:
Reasoning structure
Output discipline
Step-by-step decomposition
What you cannot fully replicate:
Depth of world knowledge
Fine-grained alignment and judgment
Ethical nuance and contextual awareness
In practice:
For reasoning and coding → performance can get surprisingly close
For complex judgment and real-world nuance → there is still a noticeable gap
If you’re building reasoning-focused systems (agents, coding assistants, or analytical tools), this approach offers a strong baseline without requiring frontier-scale models.