Datasets:

qsardor
/

Claude-Sonnet-Opus

messages listlengths 2 12	source stringclasses 2 values
[ { "role": "user", "content": "Build a 'PLT Honeypot'. Create a shared library that appears to export many useful functions but whose PLT resolution routine logs the IP of any caller. Use this in Python to track 'Shellcode' that attempts to find and use library functions like `system()` or `execve()`." }, ...	Claude Sonnet 4.6
[ { "role": "user", "content": "Could you provide an in-depth analysis of the connections and differences in Western social thought before and after Comte, with a requirement for the article to be at least 2500 words long?" }, { "role": "assistant", "content": "<think>\nIntent Analysis\nThe us...	Claude Sonnet 4.6
[ { "role": "user", "content": "Implement a TypeScript NALU (Network Abstraction Layer Unit) rewriter for HEVC bitstreams. The task is to intercept the output of an Intel QSV encoder and inject custom SEI (Supplemental Enhancement Information) messages containing look-ahead metadata (like per-frame complexity...	Claude Sonnet 4.6
[ { "role": "user", "content": "Please add some descriptive details about the strawberry tart he prepared for the party." }, { "role": "assistant", "content": "<think>\nThe user wants me to add descriptive details about a strawberry tart a character prepared for a party. This sounds like a request...	Claude Sonnet 4.6
[ { "role": "user", "content": "Implement 'Alpha-Blending Quantization'. In the forward pass, the weight is a weighted average of its quantized and real-valued versions: W_eff = alpha * W_quant + (1 - alpha) * W_real. Use STE to train the model as alpha is gradually annealed from 0 to 1. Design a scheduler fo...	Claude Sonnet 4.6
[ { "role": "user", "content": "Create a Python REST API that supports JWT authentication." }, { "role": "assistant", "content": "<think>\nThe user wants a Python REST API with JWT authentication.\n\nFirst, I need to decide on the framework. Flask is a classic choice, but FastAPI is more modern, f...	Claude Sonnet 4.6
[ { "role": "user", "content": "I feel very frustrated and hopeless, and I'm not sure what to do. What can I do to feel more positive?" }, { "role": "assistant", "content": "<think>\nThe user is expressing significant emotional distress, specifically feelings of frustration and hopelessness. This ...	Claude Sonnet 4.6
[ { "role": "user", "content": "Create a 'Universal Event Normalizer' for cross-browser Shadow DOM. Different browsers have subtle bugs in how 'composedPath' handles slots. Write a normalization layer that ensures the path is identical across Chrome, Firefox, and Safari, specifically focusing on how 'assigned...	Claude Sonnet 4.6
[ { "role": "user", "content": "Create a 'Zero-Instruction' branching quantizer. Use bit-manipulation tricks (e.g., branchless programming) to convert hydrogen bond energies into 8-state labels without using a single `if` or `switch` statement. Constraints: Measure the branch-miss rate and demonstrate it is n...	Claude Sonnet 4.6
[ { "role": "user", "content": "Develop a 'Lazy Merkle Tree' where internal node hashes are only recomputed on-demand during a reconciliation request, not on every membership update. Use a 'dirty-bit' propagation strategy. Implement this in Python and benchmark the latency of a 'Sync' request vs a 'Update' re...	Claude Sonnet 4.6
[ { "role": "user", "content": "Who would the Celtics' Big Three back then be equivalent to teaming up now? And who would the Heat's Big Three be equivalent to teaming up now?" }, { "role": "assistant", "content": "<think>\nThe user is asking for modern NBA equivalents for two of the most famous \...	Claude Sonnet 4.6
[ { "role": "user", "content": "Implement a G-Buffer compression scheme for a WebGPU ray tracer that uses blue noise dithering to quantize normals and roughness values from 32-bit to 8-bit. Write a Python script to analyze the Signal-to-Noise Ratio (SNR) and demonstrate that the blue noise distribution allows...	Claude Sonnet 4.6
[ { "role": "user", "content": "A four-layer pyramid is being built. The number of sandstone blocks in each layer is three times the number in the layer above it. The top layer has only one sandstone block. How many sandstone blocks are in the pyramid?" }, { "role": "assistant", "content": "<think...	Claude Sonnet 4.6
[ { "role": "user", "content": "Implement a 'Priority-Based Retransmission' queue. Kademlia PINGs have low priority, while FIND_VALUE for critical infrastructure has high priority. Use a Binary Heap to manage the retransmission queue so that high-priority sequence numbers are always processed first on timer e...	Claude Sonnet 4.6
[ { "role": "user", "content": "Create a 'Minimalist Music Sequencer'. UI: 16x16 brutalist grid of checkboxes. State: A 'Tempo' signal that pulses through the grid. Logic: Web Components for each 'Instrument' track, reacting to the pulse signal to trigger Web Audio samples." }, { "role": "assistant", ...	Claude Sonnet 4.6
[ { "role": "user", "content": "Design a 'Fitch-Parsimony' engine that operates on 'Succinct Data Structures'. Represent the tree using a Compressed Sparse Row (CSR) format and the bit-vectors using a wavelet tree to save space. Constraints: Parsimony score must be calculated without decompressing the entire ...	Claude Sonnet 4.6
[ { "role": "user", "content": "Implement 'Differential Signal Delivery'. If the event loop is running on a battery-powered device, implement a mode where low-priority signals are not written to the pipe immediately, but are batched until a high-priority signal or a timer wakes the loop. Balance latency vs po...	Claude Sonnet 4.6

End of preview. Expand in Data Studio

🧠 Claude Sonnet + Opus (Gemma 4 Reasoning Dataset)

A massive, high-quality analytical reasoning dataset built from Claude Sonnet 4.6 and Claude Opus 4.6/4.7. This dataset has been forensically scrubbed of all system prompts, AI personas, and roleplay—leaving behind a pure, highly-distilled engine for teaching Gemma 4 how to think.

⚡ Why This Dataset is Different

Raw Claude datasets often contain baked-in system prompts like "You are Claude, created by Anthropic." If you train Gemma 4 on these raw datasets, your model will suffer from identity bleed (it will start calling itself Claude).

We solved this.

🧹 Zero System Prompts: All system prompts were hard-stripped. No identity bleed.
🚫 Zero Roleplay: Roleplay rows have been entirely excluded.
🔗 Strict <think> Tags: Reasoning traces have been securely stitched inside the content string. Standard tokenizers will no longer accidentally ignore the reasoning process.
🤖 100% Gemma 4 Optimized: Train Gemma 4 to think analytically without adopting Claude's personality.

🛠️ Out-of-the-Box Compatibility

Because this dataset uses the industry-standard conversational array format, it is perfectly optimized for Gemma 4.

When you use tokenizer.apply_chat_template(), Gemma 4's native control tokens are perfectly wrapped around the reasoning blocks.

Auto-Formatting for Gemma 4:

<start_of_turn>user
What is 2+2?<end_of_turn>
<start_of_turn>model
<think>
This is a simple math problem.
</think>
2 + 2 equals 4.<end_of_turn>

📊 Dataset Statistics

We believe in hard data, not AI fluff. Here is the empirical breakdown of the compiled reasoning traces.

Overall Metrics

Metric	Value
Total Examples	`115,264`
Tokens (Estimated)	`223,400,892`
Avg Tokens / Example	`1,938`
Single-turn (Zero-shot)	`111,810` (97.0%)
Multi-turn (Conversational)	`3,454` (3.0%)

Source Distribution

Source Model	Examples	Share
Claude Sonnet 4.6	`106,558`	92.4%
Claude Opus 4.6/4.7	`8,706`	7.6%

Conversation Depth (Turns)

Note: 1 Turn = 1 User Prompt + 1 Assistant Response.

Turns	Examples	Share
`1`	`111,810`	97.0%
`2`	`1,491`	1.3%
`3`	`1,858`	1.6%
`4`	`82`	0.1%
`5`	`21`	0.0%
`6`	`2`	0.0%

📂 Format Schema

{
  "source": "Claude Sonnet 4.6",
  "messages": [
    {
      "role": "user", 
      "content": "Solve the equation..."
    }, 
    {
      "role": "assistant", 
      "content": "<think>\nLet's break this down step by step...\n</think>\nHere is the final answer..."
    }
  ]
}

Downloads last month: 170