omar-sharif03/loq-triples

LoQ Triples

Supervision for training a question generator for information extraction: given a document and an argument role, the set of questions that best extracts that role's arguments from that document.

The questions are not hand-written. Each is the output of an optimization loop that drafts candidate questions, extracts arguments with them, scores the extraction against ground truth, and rewrites from that feedback -- keeping whichever iteration scored best.

These are the exact rows the released question generators were fine-tuned on, not a resample: both qwen3-8b-mix-high-v1 and qwen3-4b-mix-high-v1 name this data in their run configs and encoded every row with none skipped.

Layout

Flat files, three per source corpus, so you can take only the domain you need.

<dataset>-train.jsonl      training rows
<dataset>-val.jsonl        held-out rows
<dataset>-metadata.jsonl   full optimization trace, joins on `id`

dataset	train	validation	total	metadata rows
CaseReportBench	506	114	620	620
PHEE	3,989	1,011	5,000	5,000
DiscourseEE	1,785	415	2,200	2,200
MACCROBAT	3,976	1,024	5,000	5,000
total	10,256	2,564	12,820	12,820

Training rows

field	type	description
id	string	<dataset>:<serial>, unique across the release
dataset	string	source corpus
role	string	the argument role to extract
document	string	the text the role is extracted from
questions	list[string]	the label -- the optimized questions
ground_truth_arguments	list[string]	gold arguments; what the loop scored against

{
  "id": "PHEE:train-424",
  "dataset": "PHEE",
  "role": "subject-disorder",
  "document": "Two patients treated for chronic myeloid leukemia with high doses of CCNU (1100 mg/m2 and 1240 mg/m2, respectively) developed a fatal pulmonary fibrosis.",
  "questions": [
    "What category of disease are the patients being treated for?"
  ],
  "ground_truth_arguments": [
    "chronic myeloid leukemia"
  ]
}

Train on (role, document) -> questions. The arguments are carried so each row is self-describing, not to be predicted.

Metadata rows

<dataset>-metadata.jsonl holds everything the training file omits, joined on id:

field	description
id, dataset, split	join key and which split the row is in
metadata.best_score	F1 of the winning iteration
metadata.best_iteration_number	which iteration won (0 = first draft)
metadata.total_iterations	how many refinements ran
metadata.best_trace	the winning iteration in full
metadata.trace_history	every iteration: questions, raw and parsed predictions, exact/relaxed/complex match pairs, remaining arguments, scores, leakage check
metadata.optimized_loqa_args	arguments the winning questions extracted
metadata.processed_gt_args	normalized gold arguments
metadata.document_id	the corpus's own document id (repeats across roles)
corpus fields	event, trigger, argument-type, doc_id, pmcid, post, comment, wnd_id, ... as each corpus provides

Reading the files directly

Every file is JSONL: one standalone JSON object per line. No datasets install needed.

import pandas as pd

train = pd.read_json("PHEE-train.jsonl", lines=True)
val   = pd.read_json("PHEE-val.jsonl", lines=True)

row = train.iloc[0]
row["role"]        # subject-disorder
row["questions"]   # ['What category of disease are the patients being treated for?']

Join the trace onto a training row -- both share id:

meta = pd.read_json("PHEE-metadata.jsonl", lines=True).set_index("id")

m = meta.loc[row["id"], "metadata"]
m["best_score"]             # F1 of the winning iteration
m["best_iteration_number"]  # which iteration won (0 = first draft)
m["trace_history"]          # every iteration in full
m["event"]                  # corpus fields, where the corpus has them

All four corpora at once:

DATASETS = ["CaseReportBench", "PHEE", "DiscourseEE", "MACCROBAT"]
train = pd.concat([pd.read_json(f"{d}-train.jsonl", lines=True) for d in DATASETS])

Without pandas:

import json
with open("PHEE-train.jsonl") as f:
    rows = [json.loads(line) for line in f]

Loading from the Hub

Replace REPO_ID with this dataset's id.

from datasets import load_dataset

phee = load_dataset("REPO_ID", "PHEE")            # one corpus
print(phee["train"][0])

# every corpus at once
from datasets import concatenate_datasets
all_train = concatenate_datasets([
    load_dataset("REPO_ID", name)["train"]
    for name in ["CaseReportBench", "PHEE", "DiscourseEE", "MACCROBAT"]
])

The metadata files are not exposed as configs -- trace_history nests irregularly and would break schema inference. Download and read one directly:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("REPO_ID", "PHEE-metadata.jsonl", repo_type="dataset")
meta = pd.read_json(path, lines=True).set_index("id")