Echo9Zulu/Interleave-Gemma-Distill-topk20

This dataset was used to train Shadow-Gemma-3-1B, an experimental general reasoning model trained in 9 minutes for the Google-Tunix Hackathon.

System Prompt:

You are Gemma, a thinking model who reasons through problems step by step before providing an answer.
Conduct your reasoning within a <reasoning></reasoning> block, with intermediate steps using <processing></processing> tags,
with the intermediate step inside. Continue like this until closing the </reasoning> block and providing your answer within <answer></answer>.

Validation Rules

Required structure:
    <reasoning>
        <processing>...</processing>   # At least one required
        <processing>...</processing>   # Multiple allowed
    </reasoning>
    <answer></answer>

Constraints:
    - Exactly one <reasoning> block
    - At least one <processing> block inside <reasoning>
    - Exactly one <answer> block
    - No <processing> tags outside <reasoning>
    - No <processing> tags inside <answer>
    - <reasoning> must come before <answer>

Dataset format:

    {
        "status": "accepted",
        "original_messages": [
            {"role": "system", "content": "You are Gemma..."},
            {"role": "user", "content": "Write a poem about the sea."}
        ],
        "full_prompt_text": "<bos><start_of_turn>system\nYou are Gemma...",
        "completion_text": "<reasoning><processing>First I'll consider...</processing></reasoning><answer>...</answer>",
        "logprobs_data": [
            {
                "token_index": 0,
                "top_candidates": {"<": -0.123, "The": -2.456, ...}
            },
            ...
        ],
        "attempts": 1
    }