trjxter/Gemma-4-31B-Reasoning-1000x

Gemma-4-31B-Reasoning-1000x

A 995-example reasoning distillation dataset generated with google/gemma-4-31B-it as the teacher model.

Each example is a single-turn reasoning sample formatted for supervised fine-tuning, with reasoning wrapped in <think>...</think> and the final answer after the closing tag.

Dataset repo:

trjxter/Gemma-4-31B-Reasoning-1000x

Data Structure

Each example uses the following public schema:

id
conversations
input
output
domain
meta

Each row contains:

• —conversations[0]: the user prompt.
• —conversations[1]: the model response with reasoning wrapped in <think>...</think> and the final answer after the closing tag.
• —input: prompt-only flat field for training pipelines that prefer flat prompt fields.
• —output: tagged answer-only flat field for training pipelines that prefer flat completion fields.
• —domain: task/domain category.
• —meta: lightweight per-example metadata.

Example structure:

{
  "id": "gemma4_31b_reasoning_example_id",
  "conversations": [
    {
      "from": "human",
      "value": "Solve this Python programming task..."
    },
    {
      "from": "gpt",
      "value": "<think>\nreasoning trace\n</think>\n\nfinal answer"
    }
  ],
  "input": "Solve this Python programming task...",
  "output": "<think>\nreasoning trace\n</think>\n\nfinal answer",
  "domain": "coding",
  "meta": {
    "input_tokens": 65,
    "output_tokens": 822,
    "teacher_model": "Gemma-4-31B-it",
    "teacher_model_id": "google/gemma-4-31B-it",
    "seed_source": "claudios/google-research-datasets__mbpp",
    "seed_id": "coding_example_seed_id",
    "source_id": "771",
    "source_license": "check-source-card",
    "skill": "python_problem_solving",
    "difficulty": "medium",
    "source_bucket": "coding",
    "generation_pipeline": "gemma_4_31b_reasoning_distill_1000x"
  }
}

Inference API / Generation Backend

This dataset was generated using the W&B Inference API with google/gemma-4-31B-it as the teacher model.

The generation pipeline used a teacher-student distillation style workflow:

1. 1.Build a curated seed prompt bank.
2. 2.Send each seed prompt to the teacher model through W&B Inference.
3. 3.Ask the teacher model to return structured reasoning and a final answer.
4. 4.Parse, validate, and repair the model output when possible.
5. 5.Save accepted examples in the final public SFT schema.

The final public examples use the following response format:

<think>
reasoning trace
</think>

final answer

The W&B Inference API was used only for teacher generation. The final Hugging Face dataset contains the generated examples and lightweight metadata only.

Seed Data Sources

The dataset was generated from a 1,000-prompt seed bank. The final public release contains 995 validated examples after filtering and cleanup.

Seed prompts came from a mix of public datasets and self-generated AI-engineering templates. Only the seed prompts/tasks were used from the public datasets. The original answers from public datasets were not copied into the final output field. The final answers were generated by google/gemma-4-31B-it.

Public Seed Sources

Self-Generated Seed Sources

Final Domain Distribution

The seed data was deduplicated before teacher generation to avoid repeated model calls and improve dataset diversity.

Teacher Model

Teacher model: Gemma-4-31B-it

Teacher model ID: google/gemma-4-31B-it

The teacher model was prompted to produce structured JSON with:

{
  "reasoning": "...",
  "final": "..."
}

The final dataset converter then wrapped the reasoning field inside <think>...</think> and appended the final answer after the closing tag.

Validation and Cleanup

The dataset went through multiple validation stages:

1. 1.Seed prompt validation and deduplication.
2. 2.Teacher response parsing and schema validation.
3. 3.Semantic retries for malformed JSON or failed validation.
4. 4.Raw output validation.
5. 5.Public schema export.
6. 6.Final cleanup and revalidation of train.jsonl.

Final validation checks included:

• —JSONL parseability.
• —Required public columns.
• —Conversation structure consistency.
• —input matching conversations[0].value.
• —output matching conversations[1].value.
• —<think>...</think> formatting.
• —Duplicate id detection.
• —Duplicate input detection.
• —Duplicate output detection.
• —Domain validation.

Final public row count: 995.

Intended Use

This dataset is intended for:

• —Supervised fine-tuning experiments.
• —Reasoning-format alignment.
• —Small-scale teacher-student distillation tests.
• —AI engineering workflow training examples.
• —Educational research around dataset generation and validation pipelines.

The dataset includes both chat-style and flat prompt/completion fields, so it can be used by training pipelines that expect either:

conversations

or:

input + output

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("trjxter/Gemma-4-31B-Reasoning-1000x", split="train")
print(dataset)
print(dataset[0])

To access the chat format:

example = dataset[0]
messages = example["conversations"]

To access the flat prompt/completion format:

prompt = example["input"]
completion = example["output"]

Notes

This is a small public reasoning-distillation dataset intended for experimentation and educational use.

The final file is train.jsonl, with 995 validated rows.