tcr-repro-2027/termination-control-data

Termination Control in Long Structured Generation

This dataset accompanies the study “From Hallucinated Targets to Runaway Repetition: Termination Control in Long Structured Generation.”

It contains Chinese-language relation-extraction examples with explicit candidate entity lists. The supervision conditions support controlled comparisons of how entity-boundary violations, target-block replacements, and input edits affect output repetition and termination.

The dataset consists of 13 training conditions, a shared evaluation split, and an OBR replacement manifest. Each training condition has both a structured record representation and a SWIFT chat representation. There are 28 JSONL files, totaling approximately 7.33 GB before compression. All JSONL files use UTF-8 encoding, with one JSON object per line.

Directory structure

Paths are relative to the dataset repository root.

.
├── README.md
├── cleanv2/
│   ├── train_supportclean_keep8.jsonl
│   ├── eval_supportclean_keep8.jsonl
│   └── swift_train_supportclean_keep8.jsonl
├── raw/
│   ├── train_keep4.jsonl
│   ├── train_keep4_a.jsonl
│   ├── train_keep4_ae.jsonl
│   ├── swift_train_keep4.jsonl
│   ├── swift_train_keep4_a.jsonl
│   └── swift_train_keep4_ae.jsonl
└── controlled/
    ├── train_obr.jsonl
    ├── train_obr_p15.jsonl
    ├── train_obr_p10.jsonl
    ├── train_obr_p5.jsonl
    ├── train_isc_a.jsonl
    ├── train_isc_e.jsonl
    ├── train_isc_ae.jsonl
    ├── train_benign_input.jsonl
    ├── train_generic_noise.jsonl
    ├── swift_train_obr.jsonl
    ├── swift_train_obr_p15.jsonl
    ├── swift_train_obr_p10.jsonl
    ├── swift_train_obr_p5.jsonl
    ├── swift_train_isc_a.jsonl
    ├── swift_train_isc_e.jsonl
    ├── swift_train_isc_ae.jsonl
    ├── swift_train_benign_input.jsonl
    ├── swift_train_generic_noise.jsonl
    └── obr_pair_manifest.jsonl

Dataset conditions

Directory / condition	Description
cleanv2	Cleaned supervision and the shared evaluation split.
raw/keep4	Raw supervision used in the main raw-versus-cleaned comparison.
raw/keep4_a	Raw supervision condition isolating the candidate-only violation subtype.
raw/keep4_ae	Raw supervision condition isolating the out-of-text violation subtype.
controlled/obr	Out-of-candidate block replacement at approximately 24.3%.
controlled/obr_p15	OBR at 15%.
controlled/obr_p10	OBR at 10%.
controlled/obr_p5	OBR at 5%.
controlled/isc_a	Candidate-list input edits, preserving cleaned targets.
controlled/isc_e	Source-text input edits, preserving cleaned targets.
controlled/isc_ae	Joint candidate-list and source-text edits, preserving cleaned targets.
controlled/benign_input	Benign input edits, preserving cleaned targets.
controlled/generic_noise	Generic target-label noise, preserving relation endpoints.

The OBR conditions retain the cleaned inputs and replace selected target blocks. Their replacement sets are nested: the 5% set is contained within the 10% set, which is contained within the 15% set and the full OBR set.

The three ISC conditions and the benign-input condition use validated LLM-assisted input edits. These files contain the resulting fixed examples.

Dataset sizes

An output block is one relation object in a record’s output list.

Record file or group	Records	Output blocks
cleanv2/train_supportclean_keep8.jsonl	8,854	585,029
cleanv2/eval_supportclean_keep8.jsonl	1,106	59,951
raw/train_keep4.jsonl	8,875	779,398
raw/train_keep4_a.jsonl	8,875	733,877
raw/train_keep4_ae.jsonl	8,867	630,577
Each of the nine controlled/train_*.jsonl files	8,854	585,029

Each swift_train_*.jsonl file has the same number of records as its corresponding train_*.jsonl file.

OBR condition	Replaced output blocks
obr	142,162
obr_p15	87,754
obr_p10	58,503
obr_p5	29,251

controlled/obr_pair_manifest.jsonl contains 142,162 entries, one per replacement in the full OBR condition.

Record format

The train_*.jsonl and evaluation files use the following fields:

Field	Meaning
text	Source document text.
entities_str	Candidate entity list supplied with the document.
output	A list of structured relation objects.

Each object in output contains:

Field	Meaning
source	Source entity of the relation.
target	Target entity of the relation.
relation	Relation label.
description	Textual description of the relation.

Evaluation records additionally contain key and a top-level source field identifying the corpus-source category. This top-level field is distinct from the source entity inside each relation object.

SWIFT chat format

The swift_train_*.jsonl files represent the same training examples as conversations in a messages list:

• A user message contains the extraction instructions, source text, and candidate entity list.
• An assistant message contains the serialized JSON list of target relations.

The user messages are rendered with the shared tcr/prompt_template.py builder in the accompanying code. The Qwen3 training chat template handles the empty non-thinking prefix; it should not be manually inserted into the assistant target.

Record and SWIFT files are alternative representations of the same examples and should not be combined as additional training observations.

OBR replacement manifest

controlled/obr_pair_manifest.jsonl records the fixed replacements used by the OBR conditions. Its entries include record and block locations, endpoint information, and matching diagnostics such as token-length differences.

The manifest supports inspection of the replacement construction and reconstruction of the nested OBR dose conditions.

Usage

The default Hugging Face configuration, cleanv2, selects:

• train: cleanv2/train_supportclean_keep8.jsonl
• test: cleanv2/eval_supportclean_keep8.jsonl

The test split is the shared evaluation set used across training conditions. Other supervision conditions, SWIFT files, and the replacement manifest are organized as separate repository files.

For reproduction, download the repository contents into the dataset directory used by the accompanying code, preserving the three top-level directories:

datasets/
├── cleanv2/
├── raw/
└── controlled/

Use the supervision file corresponding to the intended experimental condition. Several conditions share source documents or targets by design; they represent experimental alternatives rather than independent datasets to concatenate.