ENERZAiKR/huibeom_agentic-selfgen-ko-v1

agentic-selfgen-ko-v1

24,515 Korean agentic trajectories, generated inside a live executable environment and kept only when every turn passed a deterministic state check. Nothing here is imagined: a user-simulator LLM speaks Korean, the agent LLM calls tools, the tools actually execute against AppWorld's app backends (SQLite state), observations come from the environment — never from a model — and after each user turn a runner-side check verifies the database state (or the reply's facts) against the goal. A turn that fails is retried once with the environment intact; if it still fails, the whole episode is dropped. What remains is behaviour that verifiably worked.

The corpus now has two complementary parts:

part	rows	shape	config
single-app	5,835	deep single-domain episodes, 6–9 user sub-goals (mean 7.49)	single_app
cross-app	18,680	two apps woven together: read a value in app A, act on it in app B (mean 2.81 user turns)	cross_app
total	24,515		default

Load everything with the default config, or either part on its own:

from datasets import load_dataset
ds  = load_dataset("HBKenerzai/agentic-selfgen-ko-v1", split="train")                      # all 24,515
sgl = load_dataset("HBKenerzai/agentic-selfgen-ko-v1", "single_app", split="train")        # 5,835
crs = load_dataset("HBKenerzai/agentic-selfgen-ko-v1", "cross_app",  split="train")        # 18,680

Part 1 — single-app (5,835)

Deep episodes inside one app, 6–9 user sub-goals planned one at a time from the live environment state, spoken by a persona-conditioned Korean user simulator (존댓말/반말 mix across episodes). The agent discovers credentials itself (supervisor app → profile/passwords → login; phone uses the phone number as username), so realistic auth flows — including 401-and-recover — appear naturally.

file	rows	domain	verified by
single-simple_note.jsonl	2,084	notes: create, append, pin, delete, content/count Q&A	note records and contents
single-todoist.jsonl	1,673	projects/tasks: create, complete, delete, count	task & project records, completion flags
single-phone.jsonl	1,365	SMS to real in-world contacts, contacts, alarms, number Q&A	sent-message / contact / alarm records
single-venmo.jsonl	713	money transfers, payment requests, balance Q&A	transaction record (receiver name + amount) and exact balance delta

Part 2 — cross-app (18,680)

Each episode weaves two apps: the agent reads a concrete value in one app (a contact's number, a balance, a task count, a note's items) and must use it in another — e.g. look up a task that says "send Gina 7.5 dollars", then actually send 7.5 dollars in the money app. The runner checks both apps' state deterministically (the note now contains the number; the transaction exists with the right receiver and amount and the balance moved by exactly that much).

file	rows	pattern
cross-todoist_venmo.jsonl	3,998	a task "send N dollars to X" → an actual money transfer to X
cross-venmo_note.jsonl	3,994	read the balance → record it in a note
cross-todoist_note.jsonl	3,882	count open tasks → write the count into a note
cross-note_todoist.jsonl	3,437	read a note's items → create the matching Todoist tasks
cross-phone_note.jsonl	3,369	look up a contact's number → save it in a note

Two things the single-app part does not have:

• Pre-login variety. 40% of cross episodes start already authenticated — the relevant access_tokens are provided in the system prompt and the agent calls the feature functions directly, instead of repeating the full supervisor login dance every time. The other 60% perform the login flow. (meta.prelogin marks which.)
• De-duplicated sub-goals. Repeated identical user utterances are blocked at generation (seen_utt), so no episode asks the same thing twice.

What an episode looks like

• Every assistant round carries Korean reasoning in reasoning_content (prefill-enforced, quoted-content-aware style gate), alongside content and tool_calls. Reasoning coverage is 100%.
• Tool catalog per episode is distractor-augmented (supervisor + the episode's apps' functions plus unrelated functions, shuffled), so the model must select the right tool, not just fill arguments.
• Natural-language surfaces (user messages, reasoning, replies, user-composed argument strings) are Korean; the API layer (function names, parameter names, enum values, ids, tool observations) stays English, matching real deployments.

Schema

field	type	description
id	str	kotraj_aw-{app}-… (single) or kotraj_xaw-{scenario}-… (cross)
source	str	kotraj_appworld_{app} or kotraj_appworld_cross_{scenario}
tools	str (JSON)	OpenAI-style function catalog, distractor-augmented, English
messages	str (JSON)	chat messages; assistant messages carry reasoning_content (Korean) alongside content and tool_calls; tool messages are the environment's real JSON observations
think_source	str	selfgen_ko (single) / selfgen_ko_cross (cross)
meta	dict	persona, world_task, apps, scenario/primary, prelogin, goal/sub-goal counts, kinds

Method

• Environment: AppWorld (9 apps / 457 APIs), deterministic SQLite state checks.
• User simulator + agent: Qwen3.8-27B-FP8. Korean reasoning is prefilled and style-gated; one retry per failed turn, whole-episode drop on repeated failure.
• No LLM sits in the pass/fail decision — every kept row is a trajectory whose effect on the environment was verified by reading the app databases back.

Access

This dataset is gated: access is granted on request after manual review.