ShopTrajQA

Shopping-agent trajectory question-answering data. The dataset covers a full training + evaluation stack for building agents that reason over user shopping browsing trajectories and answer questions about them (prices, actions, products, etc.).

Contents

Split	Path	Rows	Description
SFT (train)	sft/public_shopping_multi-turn_clean.parquet	762	Multi-turn supervised fine-tuning conversations. Each row has messages (chat turns with a Python code_interpreter reasoning assistant) and tools.
RLVR (train)	rlvr/public_playground_rlvr_clean_final/	10,822	Reinforcement-learning-with-verifiable-rewards data in HuggingFace datasets on-disk (Arrow) format. Columns: data_source, prompt, ability, reward_model, extra_info. Each item is a retrieval task with a ground-truth answer, qid, reference sql, and a browsing trajectory.
Eval test	test/merged_test_32k.jsonl	100	Evaluation test set (32k-context variant).
Eval test	test/merged_test_full.jsonl	100	Evaluation test set (full variant).

Formats

SFT (sft/)

Parquet with columns:

• messages: list of {role, content} turns.
• tools: tool schema available to the assistant (a code_interpreter).

RLVR (rlvr/)

A HuggingFace datasets directory saved with datasets.save_to_disk. Load with:

from datasets import load_from_disk
ds = load_from_disk("rlvr/public_playground_rlvr_clean_final")["train"]

Fields: data_source, prompt, ability, reward_model ({ground_truth, qid, sql}), extra_info ({ground_truth, qid, sql, trajectory}).

Eval test (test/)

JSON Lines, one example per line, with fields: qid, trajectory, question, reference_answer, sql_query, difficulty, evidence_depth, has_ground_truth.

Usage

# SFT
import pandas as pd
sft = pd.read_parquet("sft/public_shopping_multi-turn_clean.parquet")

# RLVR
from datasets import load_from_disk
rlvr = load_from_disk("rlvr/public_playground_rlvr_clean_final")["train"]

# Eval
import json
test = [json.loads(l) for l in open("test/merged_test_full.jsonl")]