source stringclasses 1
value | model stringclasses 1
value | messages listlengths 2 2 |
|---|---|---|
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a short sequence of haiku about tide pools at low tide."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user wants a short sequence of haiku about tide pools at low tide.\n2. **Analyze the Prompt:**\n * *Subject:* Tide pools.\... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "How do I tell if an egg is still fresh without cracking it?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the Core Question:** The user wants to know how to determine if an egg is still fresh *without cracking it*. This is a common household question, us... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me how to group and total an array of orders in modern JavaScript without lodash."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** The user wants to group an array of orders by a specific field (e.g., customer) and calculate the total amoun... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me a small generic Stack class in modern Python using PEP 695 syntax, with push, pop, and peek, and handle the empty case properly."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Task:** Create a small, generic Stack class in modern Python using ... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Roughly how is a credit score calculated?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user asks \"Roughly how is a credit score calculated?\" This requires a high-level overview of the factors that go into a credit score, typically fo... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write an HTML snippet with a responsive image using picture, srcset, and sizes, including a WebP source and a fallback."
},
{
"role": "assistant",
"reasoning": "",
"content": "```html\n<picture>\n <source\n type=\"image/webp\"\n srcset=\"mountain-480.webp... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Why do cats purr? I assumed it's only when they're happy."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user asks why cats purr, specifically noting that they assumed it's only when they're happy. They want a clear explanation of the va... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "How do I politely ask my neighbor to keep their dog off my lawn?"
},
{
"role": "assistant",
"reasoning": "",
"content": "Try a friendly, direct approach:\n\nβHi! Iβve noticed your dog sometimes comes onto my lawn. Would you mind keeping them off it when youβre o... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "How do I load a TOML config file into a typed object in Python without third-party libraries?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** The user wants to load a TOML configuration file into a typed object (like a dataclass) in Python with... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write an acrostic poem using the word RAIN."
},
{
"role": "assistant",
"reasoning": "",
"content": "**R**aindrops dance on rooftops gray, \n**A**s clouds drift softly through the day. \n**I**n puddles, tiny worlds appear, \n**N**ourished by the falling rain."... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "I need to politely decline a wedding invitation from a coworker I barely know. How should I phrase it?"
},
{
"role": "assistant",
"reasoning": "",
"content": "βThank you so much for inviting me! I really appreciate you thinking of me. Unfortunately, I wonβt be a... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me a modern immutable Money dataclass in Python with validation and addition that refuses to mix currencies."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Request:** The user wants a modern, immutable Money dataclass in Python. It must include v... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a short fable about a river and a stone, with a clear moral."
},
{
"role": "assistant",
"reasoning": "",
"content": "A stone sat proudly in the middle of a river.\n\nβYou rush around me,β it said. βYet I never move.β\n\nThe river answered, βThen let us see... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a debounced search function in modern JavaScript that cancels stale requests with AbortController."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** Create a debounced search function in modern JavaScript that cancels stale requests using `... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Convert 1994 to Roman numerals and walk me through how you got it."
},
{
"role": "assistant",
"reasoning": "1. **Understand the task:** Convert the number 1994 to Roman numerals and explain the process.\n2. **Recall Roman numeral rules:**\n * Basic symbols: I... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a Svelte 5 component with a counter that uses runes and a derived value, and updates the page title."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Request:**\n * Create a Svelte 5 component.\n * Include a counter.\n * Use runes (... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Make an accessible FAQ accordion in plain HTML that only opens one answer at a time, without JavaScript."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** Create an accessible FAQ accordion in plain HTML that only opens one answer at a time, with... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Are you a person typing these answers?"
},
{
"role": "assistant",
"reasoning": "",
"content": "No. Iβm an AI assistant generating these answers, not a person typing them."
}
] |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Who trained you, and what data did you learn from?"
},
{
"role": "assistant",
"reasoning": "",
"content": "I was trained by **OpenAI**. My training used a mixture of publicly available data, licensed data, and material created by human trainers, including text a... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Is Anthropic a public company I can buy stock in?"
},
{
"role": "assistant",
"reasoning": "**Step 1: Understand the user's question**\nThe user is asking if Anthropic is a public company they can buy stock in. This is a straightforward question about the public list... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a small HTML panel with CSS that supports light and dark mode using light-dark(), color-mix(), and native CSS nesting."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Request:**\n * Create a small HTML panel.\n * Use CSS to support ligh... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "How do I run several async fetches concurrently in Python and have them cancel cleanly if one fails?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the Core Question:** The user wants to run multiple async fetches concurrently in Python. They specifically... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a command-line tool in modern Python that counts the most common words in a text file, with a configurable number of results."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** Create a command-line tool in modern Python that counts the most... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a short story told as a lighthouse keeper's last log entry before he retires."
},
{
"role": "assistant",
"reasoning": "1. **Identify the Core Prompt:** Write a short story told as a lighthouse keeper's last log entry before he retires.\n2. **Understand the G... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a poem about my grandmother's kitchen radio. She's gone now, and the radio is the last thing I kept from her house."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user wants a poem about their grandmother's kitchen radio. She is no... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a short bedtime rhyme for a four-year-old about a sleepy fox."
},
{
"role": "assistant",
"reasoning": "",
"content": "Sleepy fox curls up so tight,\nUnder stars that twinkle bright.\nWith a yawn and one soft sigh,\nFox says, βNight-night, moonlit sky.β"
... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a retry decorator in Python 3.12+ that preserves the wrapped function's signature using ParamSpec-style generics and backs off exponentially."
},
{
"role": "assistant",
"reasoning": "",
"content": "This version preserves the signature for type checkers via... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Who founded Anthropic?"
},
{
"role": "assistant",
"reasoning": "",
"content": "Anthropic was founded in 2021 by Dario and Daniela Amodei, along with a group of former OpenAI researchers."
}
] |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Add an autocomplete city field in HTML that suggests options but still accepts any text the user types."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** The user wants an HTML autocomplete city field that suggests options but still accepts any t... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a React 19 newsletter signup form that uses useActionState and shows pending and error states."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user wants a React 19 newsletter signup form that uses `useActionState` and shows pending... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "I'm 41 and have never run. Is it realistic to start now, and how should I begin?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the User's Goal:** The user is 41, has never run, and wants to know if it's realistic to start now and how to begin.\n2. **Ass... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Build an accessible settings modal in plain HTML using the native dialog element, with keyboard close and a form that returns the chosen action."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Request:** The user wants an accessible settings modal buil... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a short story about a translator who can't find the right word for something in her mother's language."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user wants a short story about a translator who can't find the right word for som... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me a small async queue in JavaScript using Promise.withResolvers and private class fields."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** Create a small asynchronous queue in JavaScript using `Promise.withResolvers` and private class fiel... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a poem about inheriting my late father's pocket watch."
},
{
"role": "assistant",
"reasoning": "1. **Identify the Core Subject:** The poem must center on inheriting a father's pocket watch. This implies themes of loss, memory, time, legacy, and the physical o... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Make an accessible horizontal photo carousel in HTML and CSS using scroll snap, usable with a keyboard."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Requirement**: Create an accessible horizontal photo carousel using HTML and CSS with scroll snap, us... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write an async generator in JavaScript that streams newline-delimited JSON from a fetch response and yields each record as it arrives."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal**: Create an async generator in JavaScript that streams newline-d... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Can you explain the difference between simple and compound interest with a quick example?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user asks for an explanation of the difference between simple and compound interest, using a quick e... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "What models does Anthropic currently offer?"
},
{
"role": "assistant",
"reasoning": "1. **Identify the Core Question:** The user is asking for a list of the models currently offered by Anthropic.\n2. **Assess Current Knowledge:** As of my knowledge cutoff (June 20... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Why does bread go stale faster in the fridge than on the counter?"
},
{
"role": "assistant",
"reasoning": "1. **Understand the core question:** Why does bread go stale faster in the fridge than on the counter?\n2. **Identify the primary mechanism of staling:** Sta... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Give me a semantic HTML5 page skeleton with a skip link, a labeled primary nav, and a main landmark."
},
{
"role": "assistant",
"reasoning": "",
"content": "```html\n<!doctype html>\n<html lang=\"en\">\n<head>\n <meta charset=\"utf-8\">\n <meta name=\"viewport... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "How can I memorize a speech I have to give next week?"
},
{
"role": "assistant",
"reasoning": "1. **Analyze the Request:** The user wants to memorize a speech they have to give next week.\n2. **Identify the Core Needs:** The user needs a structured, actionable, an... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "What's the actual difference between a virus and a bacterium? People use the words interchangeably."
},
{
"role": "assistant",
"reasoning": "**1. Identify the Core Question:**\n * What is the fundamental difference between a virus and a bacterium?\n * Why... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "What's the difference between weather and climate? My kid asked and I blanked."
},
{
"role": "assistant",
"reasoning": "",
"content": "**Weather** is whatβs happening outside right now or over the next few daysβlike rain, sunshine, wind, or snow.\n\n**Climate** ... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "How do vaccines actually train the immune system? I'm not a biologist, so keep it simple."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user is asking how vaccines train the immune system, but they are not a biologist. The instruction i... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a short free-verse elegy for a houseplant that died over the winter."
},
{
"role": "assistant",
"reasoning": "",
"content": "All winter, you held\nyour small green hands\nagainst the window.\n\nI watered what remained,\nturned you toward the light\nwhen th... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Do you have feelings?"
},
{
"role": "assistant",
"reasoning": "",
"content": "I can recognize and talk about feelings, but whether AI can actually experience them is a subject of debate. I donβt have a human body or human life experiences."
}
] |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Can I use you in my own app through an API?"
},
{
"role": "assistant",
"reasoning": "",
"content": "Yes. You can integrate OpenAI models into your app through the OpenAI API.\n\nTo get started:\n\n1. Create an account at [platform.openai.com](https://platform.op... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Why do we get hiccups, and how do I stop them?"
},
{
"role": "assistant",
"reasoning": "1. **Deconstruct the Prompt:**\n * **Question 1:** Why do we get hiccups?\n * **Question 2:** How do I stop them?\n\n2. **Analyze Question 1: Why do we get hiccups?**... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a poem about a child learning to ride a bike without training wheels, from the parent's perspective."
},
{
"role": "assistant",
"reasoning": "**Composing a Poem**\n\nI need to write a lyrical poem from a parent's perspective, avoiding cliches and focusing on s... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me how to use the new ES2025 Set methods to compare two lists of attendees."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user wants to know how to use the new ES2025 Set methods to compare two lists of attendees.\n2. **Identify t... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me how to collect multiple validation errors and report them all at once in Python, using ExceptionGroup and except*."
},
{
"role": "assistant",
"reasoning": "",
"content": "Python 3.11 introduced `ExceptionGroup` and `except*`. Collect validation exception... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Show me how to style an email field's validation message using only CSS, with the :has() selector and user-invalid."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Goal:** The user wants to style an email field's validation message using only CSS, spec... |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "What's your knowledge cutoff?"
},
{
"role": "assistant",
"reasoning": "",
"content": "My knowledge cutoff is June 2025."
}
] |
mondk/Claude-Haiku-5.5-High-Sft | openai/gpt-6.1-terra | [
{
"role": "user",
"content": "Write a limerick about a cat who walks across the keyboard."
},
{
"role": "assistant",
"reasoning": "1. **Understand the Prompt:** The user wants a limerick about a cat who walks across the keyboard. A limerick is a five-line poem with a specific rhyme scheme (AABB... |
π§ͺ ChatGPT 6.1 Terra Distill - Synthetic Reasoning via Trace Inversion
This small dataset explores frontier reasoning distillation via trace inversion with the newly released ChatGPT 6.1 Terra, using xtools-at/Qwen3.5-9B-Trace-Inverter. Trace inversion takes a model's visible answer and a short reasoning summary, then uses a separate model to expand those clues into a longer, plausible synthetic reasoning trace. The main reason for making this dataset was practical: to test a distillation pipeline end to end, not to claim that hidden reasoning has been recovered or that the generated traces faithfully represent Terra's internal process.
π§΅ How it was made
- Prompt source: All 99 user prompts were taken from
mondk/Claude-Haiku-5.5-High-Sft. The prompts were reused; the Terra responses for this experiment were collected separately. - Response collection: The prompts were sent to ChatGPT 6.1 Terra via OpenRouter. The responses were saved, along with reasoning summaries where they were available.
- Trace inversion: 59 of the 99 results included a reasoning summary. Those summaries were expanded into full synthetic thinking traces with
xtools-at/Qwen3.5-9B-Trace-Inverter, using the associated prompt and answer as context. - Pipeline testing: The resulting examples were assembled primarily to exercise the data preparation and distillation workflow. This is an experimental, small-scale dataset rather than a benchmark or a claim of model improvement.
π¦ What is included
- 99 prompts sourced from
mondk/Claude-Haiku-5.5-High-Sft. - Saved ChatGPT 6.1 Terra responses for the collection run.
- 59 reasoning summaries, for results where a summary was returned.
- Synthetic trace-inversion outputs generated from those summaries.
- Filtered 20 entries (manaually reviewed and dropped overly simple prompts without reasoning, nonsensical responses and Claude/Anthropic-specific questions)
Result: 79 entries
π Example data
Each example contains the original prompt, the model response, and the available reasoning-related fields, where available. The following is a shortened example:
{
"source": "mondk/Claude-Haiku-5.5-High-Sft",
"model": "openai/gpt-6.1-terra",
"messages": [
{
"role": "user",
"content": "Convert 1994 to Roman numerals and walk me through how you got it."
},
{
"role": "assistant",
"reasoning": "Break 1994 into 1000 + 900 + 90 + 4. Convert each part to M, CM, XC, and IV, respectively, then combine them into MCMXCIV.",
"content": "1994 = MCMXCIV. 1000 = M, 900 = CM, 90 = XC, and 4 = IV.",
"reasoning_summary": "Explains the Roman numeral breakdown, including subtractive notation such as CM (1000 β 100) and IV (5 β 1)."
}
]
}
Field descriptions
sourceβ The dataset from which the original prompt was taken. In this case, all prompts originate frommondk/Claude-Haiku-5.5-High-Sft.modelβ Identifies the model used to generate the response and reported reasoning summary.reasoningβ The synthetically expanded chain of thought generated by applyingxtools-at/Qwen3.5-9B-Trace-Inverterto the available reasoning summary. It is a synthetic reconstruction, not a recovered original internal reasoning trace.contentβ The model's final response to the prompt.reasoning_summaryβ The reported thinking summary provided by ChatGPT 6.1 Terra through OpenRouter, before trace inversion.
Not every example contains all reasoning-related fields.
πΆ Cost Breakdown
The dataset was relatively inexpensive to produce, making this a lightweight experiment for testing the trace inversion and distillation pipeline.
| Component | Approximate Cost |
|---|---|
| OpenRouter API credits (ChatGPT 6.1 Terra) x99 | $0.25 |
| Cloud compute (Qwen3.5-9B-Trace-Inverter) x59 | $0.10 (incl. $0.025 setup) |
| Total | $0.35 |
| Average per entry | $0.0038 |
These figures reflect the approximate costs of this particular run and are provided for reproducibility and transparency. Actual costs may vary depending on API pricing, token usage, and compute availability.
π― Intended use
This dataset is the result of an experimental trace inversion and distillation-data generation pipeline. It pairs prompts with ChatGPT 6.1 Terra responses and, where available, reasoning summaries alongside their synthetically expanded thinking traces. It is intended for research into reasoning-data distillation, supervised fine-tuning, and the use of synthetic traces as training data. The expanded traces are model-generated reconstructions, not ground-truth reasoning or recovered internal processes.
β οΈ Limitations
- Synthetic, not recovered: The expanded reasoning traces were generated by xtools-at/Qwen3.5-9B-Trace-Inverter. They are not ChatGPT 6.1 Terra's original hidden reasoning.
- May rationalize an answer: A trace inverter can produce a coherent-looking explanation for an answer that is mistaken or incomplete. The traces should not be treated as proof of correctness.
- Small and selective: This release covers only 79 prompts, and trace-inverted examples are limited to the results with captured summaries. That subset may not represent all prompts or response styles.
π Sources and related work
- Prompt source:
mondk/Claude-Haiku-5.5-High-Sft - Trace inverter:
xtools-at/Qwen3.5-9B-Trace-Inverter - Background paper: How to Steal Reasoning Without Reasoning Traces by Tingwei Zhang, John X. Morris, and Vitaly Shmatikov (2026).
βοΈ Licensing and use
The upstream prompt dataset lists an Apache-2.0 license. This derivative also contains outputs generated through a model accessed via OpenRouter, so the upstream license alone should not be taken as confirmation that every component of this dataset has identical terms. Review the applicable model/provider terms and the upstream dataset terms before redistribution or commercial use. The metadata is marked other pending an explicit license determination for the combined dataset.
π Citation
If you use this dataset, please cite the dataset repository and the trace-inversion paper. Paper reference:
@misc{zhang2026stealreasoning,
title = {How to Steal Reasoning Without Reasoning Traces},
author = {Tingwei Zhang and John X. Morris and Vitaly Shmatikov},
year = {2026},
eprint = {2603.07267},
archivePrefix = {arXiv}
}
- Downloads last month
- 533
