Datasetpublic

akashnaren/agent-ui-sft

Agent UI SFT Small synthetic supervised fine-tune (SFT) set for agent tool-use. It studies the public research question: what is the most efficient UI for agents to interact with applications and tools? Scope: rows are original lab fiction for a public agent-UI research question. Identifiers such as lab-w3, wf-synth-44, and tr-dom-9 are invented for the harness. Author Akash Premkumar (akashnaren) License Apache-2.0 Hub files train.jsonl (80), test.jsonl (20)… See the full description on the dataset page: https://huggingface.co/datasets/akashnaren/agent-ui-sft.

sourceHugging Faceapache-2.0updated 3d agoView on Hugging Face

0likes46downloads

Dataset Card

Agent UI SFT

Small synthetic supervised fine-tune (SFT) set for agent tool-use. It studies the public research question: what is the most efficient UI for agents to interact with applications and tools?

Scope: rows are original lab fiction for a public agent-UI research question. Identifiers such as lab-w3, wf-synth-44, and tr-dom-9 are invented for the harness.


Author	Akash Premkumar (`akashnaren`)
License	Apache-2.0
Hub files	`train.jsonl` (80), `test.jsonl` (20)
Mirror	Kaggle: akashpnaren/agent-ui-sft
Related	agent-ui-human, agent-ui-efficiency-scores, agent-ui-mode-pairs, ui-mode-router, Space

Schema

Each JSONL line is one object:

field	type	meaning
`id`	string	Stable id, unique across splits (e.g. `aui-cli-001`, `aui-dom-014`)
`messages`	list[object]	OpenAI-style chat: `system`, `user`, then `assistant` (often with `tool_calls`), `tool` results, final `assistant` answer
`tools`	list[object]	JSON-schema tool definitions available for that example
`ui_mode`	string enum	`cli` \	`structured_api` \	`dom_click` \	`form`
`task_type`	string	Short label (e.g. `workflow_signal`, `telemetry_diagnose`, `form_fill`, `ui_efficiency_score`)

Split balance (checkable on Hub)

split	rows	per `ui_mode`
train	80	20 × each of 4 modes
test	20	5 × each of 4 modes

Tools by `ui_mode` (as published)

`ui_mode`	tools in traces
`cli`	`run_cli`, `score_ui_trace`
`structured_api`	`call_json_api`, `workflow_op`
`dom_click`	`list_interactive`, `click_node`, `read_region`
`form`	`get_form_schema`, `patch_form`, `submit_form`

Observed message-list lengths on the published 100 rows (load the JSONL to reproduce): CLI and structured API averages ~5 messages; DOM and form averages ~7–8 (more hops for the same class of lab job).

How to load

python

from datasets import load_dataset

ds = load_dataset("akashnaren/agent-ui-sft")
print(ds)
print(ds["train"][0]["id"], ds["train"][0]["ui_mode"])

Local files (after huggingface-cli download or cloning the dataset repo):

python

from datasets import load_dataset

ds = load_dataset("json", data_files={
    "train": "train.jsonl",
    "test": "test.jsonl",
})

Pandas / Polars:

python

import pandas as pd
train = pd.read_json("train.jsonl", lines=True)
print(train["ui_mode"].value_counts())

Example row (abbreviated)

From published train.jsonl, id aui-cli-001 (ui_mode=cli, task_type=ui_efficiency_score):

—user: score lab trace tr-ui-104 for tokens / turns / recoveries (CLI vs form bakeoff).
—tool result (fiction): {"trace_id":"tr-ui-104","tokens":1840,"turns":6,"recoveries":1,"ui_mode":"form","task":"restart_worker"}
—final assistant: notes that form path is expensive vs one-shot CLI for that lab job.

Numbers inside tool results are staged lab fiction, not measurements from a live cluster.

Intended use

—Smoke LoRA / SFT on a small open model with OpenAI-style messages + tools.
—Teach routing among cli, structured_api, dom_click, and form.
—Lab exercises: Temporal-like workflow ops, diagnostics-style reads, DOM selector recovery, form validation.

Not intended as: a general tool-use corpus, a production eval suite, or a rehost of Moonshot/Kimi weights or proprietary traces.

Limitations

—100 rows total — enough for a smoke LoRA, not a general agent corpus.
—Synthetic English only.
—Workflow / metrics / token counts in tool payloads are staged, not live-cluster measurements.
—No dedicated safety / prompt-injection curriculum beyond ordinary lab refusals to invent PII or webhooks.

Agent UI SFT

Schema

Split balance (checkable on Hub)

Tools by ui_mode (as published)

How to load

Example row (abbreviated)

Intended use

Limitations

Links

Tools by `ui_mode` (as published)