DialogStudio: Towards Richest and Most Diverse Unified Dataset Collection for Conversational AI
Paper • 2307.10172 • Published • 13
id stringclasses 10
values | num_train int64 8.19k 8.19k | train_data listlengths 8.19k 8.19k | num_dev int64 828 828 | dev_data listlengths 828 828 | num_test int64 1k 1k | test_data listlengths 1k 1k |
|---|---|---|---|---|---|---|
abcd | 8,192 | [[{"role":"assistant","content":"Hi, how can I help you today?","tool_calls":null},{"role":"user","c(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hello, how can i help you","tool_calls":null},{"role":"user","conte(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Hello, how can I help you today?","tool_calls":null},{"role":"user"(...TRUNCATED) |
abcd-1 | 8,192 | [[{"role":"assistant","content":"hi how can I help?","tool_calls":null},{"role":"user","content":"i (...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hi, how can I help you?","tool_calls":null},{"role":"user","content(...TRUNCATED) | 1,001 | [[{"role":"user","content":"Hi so many of your items are always out of stock. It's very frustrating(...TRUNCATED) |
abcd-2 | 8,192 | [[{"role":"assistant","content":"Hello. How can I help you today?","tool_calls":null},{"role":"user"(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hello, how can i help you today","tool_calls":null},{"role":"user",(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Thanks for contacting AcmeBrands! How can I help you today?","tool_(...TRUNCATED) |
abcd-3 | 8,192 | [[{"role":"assistant","content":"Hi! Thanks for contacting Acme support. How can I help you today?",(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hi! how can I help you today?","tool_calls":null},{"role":"user","c(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Hello. Thanks for contacting Acme customer support, how may I help (...TRUNCATED) |
abcd-4 | 8,192 | [[{"role":"assistant","content":"Hello, how can i help?","tool_calls":null},{"role":"user","content"(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hi! How can I help you?","tool_calls":null},{"role":"user","content(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Hello, thank you for contacting AcmeBrands. How can I help today?"(...TRUNCATED) |
abcd-5 | 8,192 | [[{"role":"assistant","content":"hello","tool_calls":null},{"role":"assistant","content":"how can i (...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hello, how can I help you today?","tool_calls":null},{"role":"user"(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Hello!","tool_calls":null},{"role":"assistant","content":"how can i(...TRUNCATED) |
abcd-6 | 8,192 | [[{"role":"assistant","content":"Hello. How may I help you today?","tool_calls":null},{"role":"user"(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hello how can i help you","tool_calls":null},{"role":"user","conten(...TRUNCATED) | 1,001 | [[{"role":"user","content":"Hi","tool_calls":null},{"role":"assistant","content":"Hi, how can I help(...TRUNCATED) |
abcd-7 | 8,192 | [[{"role":"assistant","content":"Hello, how may I help you today?","tool_calls":null},{"role":"user"(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hello, how can I help you?","tool_calls":null},{"role":"user","cont(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Hi! How can I help you today?","tool_calls":null},{"role":"user","c(...TRUNCATED) |
abcd-8 | 8,192 | [[{"role":"user","content":"Hi","tool_calls":null},{"role":"user","content":"I have a request for yo(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Good morning, how can I help you today?","tool_calls":null},{"role"(...TRUNCATED) | 1,001 | [[{"role":"assistant","content":"Hello, how can I help you?","tool_calls":null},{"role":"user","cont(...TRUNCATED) |
abcd-9 | 8,192 | [[{"role":"assistant","content":"Hi there,","tool_calls":null},{"role":"user","content":"Hi I have a(...TRUNCATED) | 828 | [[{"role":"assistant","content":"Hello, how may I help you?","tool_calls":null},{"role":"user","cont(...TRUNCATED) | 1,001 | [[{"role":"user","content":"Hi","tool_calls":null},{"role":"assistant","content":"Hi, how may I assi(...TRUNCATED) |
ABCDBench
10 episodic rows with train_data, dev_data, and test_data
conversation lists, plus their counts. The physical Hugging Face split is
train; the three disjoint benchmark splits live inside each row. Every row
contains the same conversations in each split, in a different shuffled order.
Rows measure ordering variation, not independent samples of conversations.
Source: Salesforce/DialogStudio, joined by original conversation ID to ABCD v1.1. DialogStudio omits action turns; the original release restores the complete chronological message/action sequence, including final customer messages. Source membership is checked exactly in each split before conversion.
First filter conversations exceeding 1024 tokens under the full Horizon
template with meta-llama/Llama-3.2-1B-Instruct revision 9213176726f574b556790deb65791e0c5aa438b6, default system
metadata dated 26 Jul 2024, and no tools list. This includes headers, tool
calls/responses, and end-of-turn tokens. Excluded from original train/dev/test:
15/3/3.
No truncation is applied.
Using Python Random seed 42, shuffle eligible dev, move
173 conversations into the filtered original train,
then shuffle each resulting split. Counts: train 8192,
dev 828, test 1001 in every row. Row zero (abcd)
preserves the original single-row dataset order. Each additional row abcd-i
independently shuffles each row-zero split with Python Random string seed
"42:i:split". Membership and the dev-to-train transfer are fixed
across rows. audit.json records every row's split IDs and shuffle seeds;
top-level split IDs refer to row zero.
Maximum rendered lengths: train 1014,
dev 1014, test 1016 tokens.
Original customer/agent text becomes user/assistant messages. Each action
becomes an assistant tool_calls entry followed by its original response as
a tool message. Function names and ordered argument values come from ABCD's
action annotations; arguments are serialized as {"values": [...]} to preserve
positional values without inventing names for polymorphic slots. These messages
match outer-loop-jax's Horizon chat template, including one tool call per message.
No intent labels, scenario, guidelines, system message, or tool catalog is stored.
The evaluator adds the standard Horizon system metadata and scores the complete
assistant tool-call spans separately from ordinary assistant dialogue.
Rebuild with uv run --no-sync python data_preparation/benchmarks/abcd_bench.py in
outer-loop-jax. Upstream data/code are MIT licensed; see LICENSE.
Please cite Chen et al. (2021), ABCD
and Zhang et al. (2024), DialogStudio.
- Downloads last month
- 190