Datasets:

atenareply
/

orbital-mining-corpus-seeded

license: apache-2.0
pretty_name: Orbital Mining Corpus — Seeded
language:
  - en
task_categories:
  - text-generation
tags:
  - domain-adaptation
  - synthetic
  - fictional
  - grounded-generation
size_categories:
  - n<1K
dataset_info:
  features:
    - name: text
      dtype: string
    - name: doc_type
      dtype: string
    - name: global_index
      dtype: int64
  splits:
    - name: train
      num_bytes: 40676203
      num_examples: 864
  download_size: 19694579
  dataset_size: 40676203
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Orbital Mining Corpus — Seeded

What's inside — The 848 Orbital Mining Corporation technical documents plus 16 new synthetic OMC docs (incident reports, ops logs, handovers, test reports, maintenance procedures).

Where it comes from — orbital-mining-corpus, extended by CROSS-SEEDING: teacher-written docs (Scaleway, gpt-oss-120b) anchored to real telemetry values from the Mars facts table.

How it was built / modified — Strict grounding — only candidates whose telemetric values appear VERBATIM in the source facts survived the filter (16/150 candidates), with a deterministic semantic judge as backstop and a hard generation cap against model collapse.

Overview

Role: Data preparation — cross-seeded OMC documents
Lineage: orbital-mining-corpus + telemetry facts (curate_logs) → cross_seed_corpus.py → this → noval-corp-training-corpus-lfm2

Schema

Column	Meaning
`text`	the document
`doc_type`	OMC document type
`global_index`	stable row id

Splits

Split	Rows	Note
train	864	848 base OMC docs + 16 grounded seeded docs

Used by

OMC side (~84% by token) of noval-corp-training-corpus-lfm2
Track A analysis-QA source (build_agentic_sft.py)

Limitations

Seeded yield was low by design (strict grounding): 16/150 candidate docs survived (~+1.9% enrichment).

Part of the noval-corp fictional-domain training family (Orbital Mining Corporation / Mars Express; Asterion round for Gemma-4). Card generated by noval-corp/scripts/gen_dataset_cards.py — see noval-corp/NOTE.md for the full project log.