zestehl/supersymmetry-v1

Supersymmetry v1

96,990 documents across 21 types — the first and only ML-ready dataset for GregTech/Supersymmetry modded Minecraft.

Overview

This dataset is a comprehensive natural-language knowledge base extracted from the Supersymmetry (SuSy) modpack, a GregTech-based Minecraft modpack. Every document was generated by a deterministic pipeline from structured game data and rendered into fluent natural language by DeepSeek V4 models.

Uniqueness: No comparable dataset exists for GregTech, Supersymmetry, or any Minecraft modpack at this scale. The corpus covers materials, recipes, machines, quests, fluids, crafting, ore deposits, cross-modpack material mapping, tool behaviors, and multiblock descriptions — all rendered into embedding-optimized paragraphs.

Dataset Structure

published/supersymmetry-v1/
  data/
    train-00000.parquet    (87,281 docs, 23.2 MB)
    test-00000.parquet     (9,709 docs, 2.6 MB)
  README.md

Schema

Field	Type	Description
id	string	Document ID (e.g., material:gregtech.material.copper)
type	string	Document type (21 types, see below)
text	string	Rendered natural language paragraph
source_text	string	Original structured field data from game files
source_model	string	Rendering model (deepseek-v4-flash or deepseek-v4-pro)
track	string	Provenance track (phase2, A, B, C, E)
tokens_in	int64	Prompt tokens sent to render model
tokens_out	int64	Completion tokens received from render model

Document Types (21)

Type	Count	Track	Description
recipe	41,900	phase2	Machine recipes with I/O, tiers, energy
item	26,984	phase2	Game items with material bindings and metadata
crafting_recipe	12,738	A	Vanilla crafting table recipes (shaped/shapeless)
groovy_recipe	4,106	phase2	Groovy-defined custom machine recipes
material	3,412	phase2	Materials with properties, composition, forms, state
fluid	2,426	phase2	Fluid definitions with usage statistics
machine	1,644	phase2	Machine definitions with recipemaps
quest_step	1,071	B	BetterQuesting quest steps (name, desc, prereqs, tasks)
cross_modpack_material	737	E	GregTech materials mapped across all major packs
smelting_recipe	635	A	Vanilla furnace recipes
process_chain	500	phase2	Top-500 material production/consumption chains
wiki_concept	220	phase2	Wiki concept documents
ore_chain	156	A	Ore → mineral → metal processing chains
recipemap	155	phase2	Machine recipemap definitions
ore_deposit	101	phase2	Worldgen ore deposit data
material_pipe	48	A	Material fluid/item pipe properties
material_wire	46	A	Material wire electrical properties
lang_description	45	C	Multiblock/cover/enchantment descriptions from lang
quest_chain	31	B	BetterQuesting quest lines as progression chains
material_tool	24	A	Material tool properties
tool_behavior	12	C	GregTech tool behavior descriptions

Train/Test Split

Stratified 90/10 split by document type. Each type appears proportionally in both splits. Random seed: 42.

Data Derivation

Source Pipeline

entity_registry.db (42 MB, 30 tables)
    ↓ SQL extraction + templates
corpus.jsonl (structured field-labeled data)
    ↓ DeepSeek V4 rendering (v4-flash + v4-pro)
rendered_corpus.jsonl (natural language paragraphs)
    ↓ BGE-base-en-v1.5 embedding (768-dim)
embedding_cache.db (pre-computed vectors)
    ↓ ChromaDB index
supersymmetry-v1 (vector search index)

Expansion Tracks

The dataset was built incrementally across five tracks:

• Phase 2 (81,448 docs): Material properties, recipes, items, fluids, machines, ore deposits, process chains, wiki concepts, recipemaps.
• Track A — DB Tables (+13,647): Crafting recipes, smelting recipes, material tools/wires/pipes, ore processing chains, and material alias enrichment.
• Track B — Quest Book (+1,102): BetterQuesting quest steps and quest lines from the SuSy modpack, resolved through the quest lang file (2,248 entries).
• Track C — Lang Mining (+57): High-signal tool behavior descriptions and multiblock language entries from the 11,428-row lang_entries table.
• Track E — Cross-Modpack (+737): GregTech material IDs mapped across GTNH, Nomifactory, Omnifactory, and Supersymmetry.

Rendering

Documents were rendered to natural language using:

• DeepSeek V4 Pro (thinking mode, $0.435/M in): 3,412 material documents
• DeepSeek V4 Flash ($0.14/M in): 93,578 non-material documents

Total inference cost: ~$0.67. Zero failures across 96,990 API calls.

Quality

• Answer relevance: 100% on 20 curated RAG queries
• Retrieval latency: 20-30ms avg (BGE-base semantic via ChromaDB)
• Generation latency: 4.5-5s avg (DeepSeek V4 Flash)
• Type coverage: All 21 document types verified present in both splits
• Zero duplicates: Deterministic ID generation verified

Usage

Load with HuggingFace Datasets:

from datasets import load_dataset
dataset = load_dataset("orthrus-research/supersymmetry-v1")

Or directly with pyarrow:

import pyarrow.parquet as pq
train = pq.read_table("data/train-00000.parquet")

For RAG: use text field for embedding (BGE-base-en-v1.5, 768-dim). For fine-tuning: use source_text for structured input and text as target output.

Citation

@dataset{supersymmetry-v1,
  title={Supersymmetry v1},
  author={Orthrus Research},
  year={2026},
  description={Natural-language knowledge base extracted from Supersymmetry modpack. 96,990 documents across 21 types.},
  url={https://huggingface.co/datasets/orthrus-research/supersymmetry-v1}
}

---

license: gpl-3.0 language: - en tags: - Supersymmetry - Gregtech - Forge - 1.12.2 - Minecraft size_categories: - 10K<n<100K