Strata-distilled

A small test dataset (~50 conversations) of synthetic chat data with reasoning traces at several thinking levels. It can be used to fine-tune smaller thinking models.

Test release. This is a proof-of-concept run of a larger generation pipeline, not a finished dataset.

Disclaimer

• Generated by a heavily quantized model. All data was produced by swift-1.5-iq3_xxs, an IQ3_XXS (roughly 3-bit) quantization. Heavy quantization lowers answer quality and reasoning reliability, so expect factual errors, flawed logic, weak code and inconsistent formatting.
• Not checked before publishing. The samples are raw model output. They were not reviewed by a human or verified for correctness, safety or bias. Do not treat the answers or the reasoning as ground truth.
• Use it to test a training pipeline. Filter and validate it before relying on it for anything serious.

How it was made

One teacher model (swift-1.5-iq3_xxs, served locally through an OpenAI-compatible API) played both sides of every conversation:

1. The user. The model was prompted to write casual, human-sounding messages. Each conversation got a random persona, writing style and message length, and a concrete topic idea brainstormed by the model, so messages read like real chat input (lowercase, slang, typos, terse or rambling).
2. The assistant. The model answered each turn, with thinking turned on or off depending on the sample's thinking level.

Conversations have 1 to 4 turns, and some start with a system prompt.

Topics: coding, MMLU-style multiple-choice questions across the 57 MMLU subjects, world knowledge, math and reasoning, writing and language, practical advice, instruction following and casual chat.

Thinking levels

Each sample has a thinking_level field:

Level	Meaning
off	No reasoning. The assistant turn begins with an empty <think> block.
low	Very short reasoning.
medium	Focused, step-by-step reasoning.
high	Long, thorough reasoning with double-checking.

Levels were requested through the prompt, and samples whose reasoning length didn't fit the requested level were discarded. They are approximate labels, not guarantees.

Format

The file is a JSON array. Each element looks like this:

{
  "id": "a1b2c3d4e5f6",
  "category": "coding",
  "subcategory": "Python",
  "thinking_level": "medium",
  "num_turns": 2,
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "<think>\n...reasoning...\n</think>\n\nfinal answer"}
  ]
}

• system messages are optional.
• Assistant turns follow the Qwen3 chat format. The reasoning is inside <think>...</think>, followed by the answer. With off, the block is empty (<think>\n\n</think>\n\n).
• messages is the standard chat format used by most fine-tuning libraries (TRL, Unsloth, Axolotl and others).

Usage

from datasets import load_dataset

ds = load_dataset("json", data_files="swift_distill_dataset.json", split="train")
print(ds[0]["messages"])

Limitations

• Tiny sample size, so it is not suitable for real fine-tuning on its own.
• Single teacher model, so its mistakes and style biases are copied into the data.
• Reasoning traces come from a heavily quantized model and may be wrong even when the final answer is right (or the reverse).
• English only.
• Simulated users are imitations of human messages, not real user data.

License

MIT, as declared in the metadata above.