README.md

metadata

language:
  - en
tags:
  - gemini
  - distillation
  - agentic
  - code-generation
  - reasoning
  - multimodal
  - instruction-following
  - synthetic
  - jsonl
  - zero-duplicates
pretty_name: Gemini 3.5 Flash Distilled Dataset (25k)
size_categories:
  - 10K<n<100K

Gemini 3.5 Flash Distilled Dataset (25k)

A 25,000-sample synthetic distilled dataset designed to replicate the core capabilities of Gemini 3.5 Flash: frontier-level agentic execution, rapid multi-step reasoning, dense context analysis, and advanced autonomous coding — all optimized for low-latency inference.

Dataset Summary

This dataset was created via template-based evolutionary synthesis with content-normalized SHA-256 deduplication. Every sample features explicit <thought>-block reasoning, structured outputs, and high-density token distributions — mirroring Gemini 3.5 Flash's thinking-level control and sub-agent execution style.

Capability	Samples	Description
Agentic Code Synthesis	7,500	Self-correcting logic, recursive algorithmic optimization, state-machine patterns, structural UI design
Dense Context Reasoning	6,000	Analytical extraction, cross-document variable tracking, root cause analysis over synthesized diagnostic documents
Multimodal Structural Mapping	5,000	Structured textual representations of audio frequencies, video frame sequences, and layout matrices
Mathematical Engine Traces	3,500	Symbolic computations, execution step validation, matrix calculations with structured JSON state traces
Systemic Execution Instructions	3,000	Complex tool calling syntax, structured JSON schema constraints, multi-turn API call trees

Dataset Structure

JSONL Format

Each line is a JSON object with three fields:

{
  "instruction": "Design a highly optimized, recursive state-machine pattern for application block 104...",
  "output": "<thought>\n1. Establish immutable state tracking boundaries...\n</thought>\n\nclass StateEngineBlock104:\n    ...",
  "metadata": {
    "category": "agentic_code_synthesis",
    "index": 104
  }
}

Data Fields

Field	Type	Description
instruction	string	The task prompt / user query
output	string	Model response with <thought> block + structured content
metadata.category	string	One of the 5 capability categories
metadata.index	integer	Sequential index within category

Features

• Structured reasoning: Every response contains an explicit <thought> block with numbered reasoning steps before the final answer.
• Zero duplicates: Content-normalized SHA-256 hashing ensures no exact or near-identical entries.
• Token-dense: No filler text — every sample contains high-density logic, code, or structured data.
• File size: 52.3 MB uncompressed (~12 MB with gzip).

Dataset Creation

Methodology

1. High-cardinality template pools: Per-category generators draw from pools of 100-1000+ variable values (components, algorithms, languages, protocols, metrics, etc.), producing millions of unique prompt combinations.
2. Content-normalized deduplication: Every generated entry is hashed after lowercasing, whitespace removal, and punctuation stripping. Duplicates are skipped.
3. Streaming output: Entries are written directly to JSONL — no in-memory accumulation.
4. Deterministic seed: random.seed(42) for reproducibility.

Deduplication Guarantee

Metric	Value
Total samples	25,000
Exact duplicates	0
Near-duplicates (normalized)	0
Unique content hashes	25,000

Requirements for Training

pip install datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files="gemini_35_flash_distilled_25k.jsonl")
print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['instruction', 'output', 'metadata'],
#         num_rows: 25000
#     })
# })

Intended Use

This dataset is intended for supervised fine-tuning (SFT) and distillation of large language models to produce efficient sub-agents with:

• Fast, structured reasoning (<thought> loops)
• Multi-step tool orchestration
• Autonomous code generation and debugging
• Dense context analysis and root cause extraction
• Multimodal structural mapping (audio-visual temporal analysis)
• Mathematical computation traces
• Strict instruction and format adherence

Considerations

• Synthetic data: All samples are programmatically generated. No real user interactions or proprietary information.
• Domain coverage: Focused on software engineering, enterprise tool use, mathematical reasoning, multimodal analysis, and structured data tasks.
• No PII: No personal or sensitive content.