TriDeHall-ECP

TriDeHall-ECP is a bilingual Chinese-English scientific preference corpus for training and evaluating scientific knowledge agents with DPO-style preference data.

The dataset accompanies the NLPCC 2026 paper TriDeHall: A Unified Tri-Level Hallucination Mitigation Framework for Multimodal Cross-Lingual Scientific Knowledge Agents.

Dataset Statistics

Split	Language	Valid preference records
full/chinese	Chinese	625,592
full/english	English	723,929
full total	Chinese + English	1,349,521
train_20k	Chinese + English	20,000

The corpus contains approximately 5.52B estimated tokens in the paper-level accounting. File-level record counts, byte sizes, and SHA-256 checksums are listed in MANIFEST.csv.

What Is an Epistemic Contrastive Pair?

TriDeHall-ECP differs from preference datasets that only contrast correct and incorrect answers.

Each record contains a user question, a chosen response, and a rejected response. The intended preference signal targets epistemic calibration in scientific explanation: the chosen response is a verified and decontextualized scientific answer whose scope is calibrated to the available evidence, while the rejected response is more source-bound, under-generalized, or epistemically miscalibrated.

The current JSONL records expose the final chosen and rejected messages used for preference learning. Record-level ECP strategy labels, policy-variant labels, bilingual alignment IDs, and page-level evidence locators are not included in this release.

Data Construction

TriDeHall-ECP is generated from scientific literature using the TriDeHall pipeline:

1. Scientific papers are parsed into text and multimodal content. Visual Transcription Anchoring (VTA) handles visual symbols before semantic interpretation.
2. A paper-level representation is constructed, followed by role-conditioned question generation.
3. TGVP generates a source-grounded answer and then performs bounded enrichment under a fact-integrity constraint.
4. FSCL verification checks source-bound or over-generalized statements and revises or filters failed generations.
5. ECP construction creates preference pairs by contrasting verified enriched answers with lower-generality or source-bound alternatives.
6. Cross-lingual filtering is used in the paper to support bilingual consistency.

Dataset Structure

Each JSONL record follows this structure:

{
  "conversations": [
    {
      "from": "human",
      "value": "..."
    }
  ],
  "chosen": {
    "from": "gpt",
    "value": "..."
  },
  "rejected": {
    "from": "gpt",
    "value": "..."
  },
  "metadata": {
    "source_file": "...",
    "language": "chinese"
  }
}

metadata/schema.json defines the released record schema.

Configurations and Splits

• full: the full uploaded corpus with chinese and english splits.
• train_20k: a fixed public 20,000-record split containing the first 10,000 Chinese records and the first 10,000 English records from this release.

train_20k is a subset of full. Therefore, summing all rows in MANIFEST.csv gives 1,369,521 file records: 1,349,521 full-corpus records plus the duplicated 20,000-record public training split.

Use train_20k for smoke tests and lightweight reproduction checks. For paper-exact reproduction, follow the split and protocol released with the accompanying code repository if an experiment uses a separate stratified or BCC-filtered sample.

Loading

from datasets import load_dataset

train = load_dataset("Airjiannan/TriDeHall-ECP", "train_20k", split="train")
zh = load_dataset("Airjiannan/TriDeHall-ECP", "full", split="chinese")
en = load_dataset("Airjiannan/TriDeHall-ECP", "full", split="english")

For large-scale processing, use streaming:

from datasets import load_dataset

zh_stream = load_dataset(
    "Airjiannan/TriDeHall-ECP",
    "full",
    split="chinese",
    streaming=True,
)

Provenance and Copyright

The corpus is constructed from published scientific literature used through compliant access channels. The released files contain generated preference records rather than redistributed full source papers.

The source_file metadata field provides the source identifier retained in this release. The released JSONL files do not include DOI, page, section, figure, table, or paragraph-level evidence locators.

Intended Use

This dataset is intended for academic research on scientific LLMs, preference optimization, epistemic calibration, hallucination mitigation, bilingual scientific QA, and reproducibility studies related to the TriDeHall framework.

It should not be treated as a universal scientific-truth dataset or as a substitute for expert verification in high-stakes scientific, medical, legal, financial, or engineering decisions.

Limitations

• The current corpus focuses on Chinese and English.
• The source domain is primarily geoscience literature.
• The construction pipeline relies partly on LLM-based generation and verification, so model-specific biases may remain.
• Bounded background enrichment may introduce statements that are not individually verified against a source passage.
• Automatic filtering does not guarantee that every preference pair is error-free.
• The public train_20k split is a fixed release split; users reproducing a paper-specific experiment should follow the exact experimental protocol released with the code.

License and Use Terms

This dataset is released under the TriDeHall-ECP Research Use Terms. The license text is provided in LICENSE.

The dataset is intended for academic research, reproducibility, and non-commercial evaluation. It contains generated scientific preference records derived from published literature and does not redistribute full source papers. Users are responsible for complying with applicable source-publication rights, institutional policies, and local laws.

Code

The accompanying reproduction code is available at: https://github.com/Airjiannan05/TriDeHall

Citation

If you use TriDeHall-ECP, please cite the accompanying NLPCC 2026 paper. BibTeX will be updated when the proceedings or preprint metadata becomes available.