sxiong/TG_LLM_output

TG-LLM Output

This dataset contains inference outputs from the two-stage TG-LLM framework on the TGQA test set. TG-LLM performs temporal reasoning in two steps:

1. Story-to-Temporal-Graph Translation (Story2TG): converts a story into a temporal graph.
2. Temporal-Graph Reasoning (TGR): reasons over the predicted temporal graph to answer temporal questions.

The original TGQA dataset is available at sxiong/TGQA, and the source code is available in the TG-LLM repository.

Models

Both LoRA adapters use meta-llama/Llama-2-13b-chat-hf as the base model.

• Story2TG: sxiong/TG-LLM-TGQA-Story2TG-Llama-2-13B-LoRA
• TGR: sxiong/TG-LLM-TGQA-TGR-Llama-2-13B-LoRA

Dataset Configurations

Configuration	Split	Rows	Description
TGQA_story_TG_trans	test	101	Stories, reference temporal graphs, prompts, and predicted temporal graphs.
TGQA_TGR	test	3,316	Temporal questions and answers generated from the Story2TG predictions.

The TGR prompts use the predicted temporal graphs produced in the first stage. The meta.TG field retains the reference temporal graph from TGQA for analysis; it is not the graph included in the TGR prompt.

Inference Configuration

• Precision: BF16
• Prompt format: plain text
• In-context learning: disabled
• Decoding: deterministic greedy decoding (do_sample=False, num_beams=1)
• Story2TG maximum new tokens: 1,024
• TGR maximum new tokens: 512

No sample reached its maximum generation length.

Data Fields

TGQA_story_TG_trans

• id: Story identifier.
• story: Input story.
• TG: Reference temporal graph from TGQA.
• prompt: Full prompt provided to the Story2TG model.
• prediction: Temporal graph generated by the Story2TG model.

TGQA_TGR

• id: Question identifier.
• story: Source story.
• question: Temporal reasoning question.
• answer: List of reference answers.
• prompt: Full TGR prompt containing the predicted temporal graph.
• prediction: Generated reasoning trace and answer.
• correct: Whether the parsed prediction exactly matches a reference answer.
• meta: Additional reference data:
  • TG: Reference temporal graph from TGQA.
  • candidates: Candidate answers.
  • external knowledge: Temporal relations and arithmetic supplied in the prompt.
  • Q-Type: TGQA question-type index.

For correct, answers are compared case-insensitively after removing spaces. For question types 2 and 3, only the leading numeric value is compared, matching the TG-LLM evaluation protocol.

Results

All 3,316 TGR predictions were successfully parsed: Exact match (macro average) -- 77.5%.

Accuracy by Task

Q-Type	Task	Correct / Total	Exact Match
0	Pairwise start-time ordering	290 / 303	95.7%
1	Five-event chronological ordering	830 / 990	83.8%
2	Event duration	168 / 221	76.0%
3	Time between event starts	238 / 292	81.5%
4	Event start time	280 / 303	92.4%
5	Same start year	289 / 303	95.4%
6	Event overlap / still happening	147 / 212	69.3%
7	Immediate predecessor / successor	210 / 504	41.7%
8	Event-duration comparison	116 / 188	61.7%

Loading the Data

After uploading this directory to a Hugging Face dataset repository, load either configuration with datasets:

from datasets import load_dataset

story_to_tg = load_dataset(
    "sxiong/TG_LLM_output",
    "TGQA_story_TG_trans",
    split="test",
)

tgr = load_dataset(
    "sxiong/TG_LLM_output",
    "TGQA_TGR",
    split="test",
)

Citation

@inproceedings{xiong2024large,
  title={Large language models can learn temporal reasoning},
  author={Xiong, Siheng and Payani, Ali and Kompella, Ramana and Fekri, Faramarz},
  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={10452--10470},
  year={2024}
}