ti_5k

Screen-grounding training cell from the gui-agent-research pipeline (datagen v2 text negatives + image-neg synthetic screenshots), packed with every screenshot it references so it trains without the AgentNet image archives.

rows	steps	trajectories	text rows	image rows	ref rows	images	size
23488	17033	4428	11744	11744	0	48176	30.3 GB

Systems: {'Windows': 5071, 'Darwin': 5020, 'Ubuntu': 13397}. Negative types: {'N-stale': 6059, 'I-D': 10779, 'N-linger': 5475, 'N-anchor': 181, 'I-D0': 965, 'N-act': 29}. State-line format: list.

Which cell to train on (read this first)

This dataset holds two renderings of the same 23,488 rows (same keys, order, steps, negatives and images). Each belongs to exactly one base model. Never mix them, and never reuse a reference (ref) file across models.

base model	train on	NOT this	format	reference log-probs
Qwen/Qwen3.5-9B	cells/ti_5k.jsonl	ti_5k_venus.jsonl	OSWorld Qwen agent (vendor tool calls)	compute with Qwen3.5-9B on ti_5k.jsonl
inclusionAI/UI-Venus-2-9B	cells/ti_5k_venus.jsonl	ti_5k.jsonl	official UI-Venus-2 Computer agent (<think>..</think><action>..</action>, 0-999 coords)	compute with UI-Venus-2-9B on ti_5k_venus.jsonl

• Feeding ti_5k.jsonl to UI-Venus-2, or ti_5k_venus.jsonl to Qwen3.5, trains the model in a format it is not evaluated in. Don't.
• Neither cell ships a ref file; compute it under the model you train (training-handoff: scripts/ref.sh --model qwen35 / --model venus).
• ti_5k_venus.jsonl was added on 2026-11-06 (final version: builder commit 0e1da0e, scroll = Sequence([Hover(box), Swipe(...)])); cells/ti_5k_venus.stats.json has its sha256, builder commit and format pins. ti_5k.jsonl and the images are unchanged.

Files

• cells/ti_5k.jsonl -- Qwen3.5-9B only. One row per (step, negative): key, tid, t, system, instruction, images (the step's screenshot window, paths relative to this dataset), messages (chat history up to the step), pos / neg (the two assistant chains; a text negative differs in the <current_state> line, an image negative has the same chain and images_neg = the window with its last screenshot replaced by the synthetic o^-), type, pivot, fmt.
• cells/ti_5k.ref.jsonl (only when ref rows > 0; otherwise run ref.py first, see below) -- the same rows with ref_pos / ref_neg / n_pos / n_neg / prompt_tok: sequence log-probs of the stock policy Qwen/Qwen3.5-9B (its chat template, the qwen_vendor message format and image preprocessing) used as the DPO / contrastive reference. These numbers are valid only when that exact model is the reference. For UI-Venus-2 train on cells/ti_5k_venus.jsonl with its own ref.
• cells/ti_5k_venus.jsonl -- UI-Venus-2-9B only. The same rows in the official UI-Venus-2 Computer format (github inclusionAI/UI-Venus @ 6d5f8ba, models/computer/computer_example.py, n_img 2, reasoning on): messages = system prompt with the task, each prior step as user ("" or "History Screenshot:" + image) + assistant <think>{state block + action prose}</think>\n<action>{Venus action}</action>, then "Current Screenshot:" + image; pos / neg chain = the continuation after the generation prompt <think>\n, i.e. {thought}\n</think>\n\n<action>...</action>. Actions map 1:1 from the AgentNet ground-truth code (Click(box=(x, y)), Type(content='...\n'), Drag, Swipe, Hotkey, ...). The thought carries the same state claims as ti_5k.jsonl, so the contrast (which claim is wrong) is identical. Same images / images_neg paths.
• cells/ti_5k_venus.stats.json -- counts, sha256, model / vendor / builder pins of the Venus cell.
• images/ -- AgentNet frames (<uuid>.png) and synthetic negatives (<tid>_<t>_<...>.png); packed as images/part-NNNNN.tar shards (fetch unpacks them in place).
• stats.json -- counts and the sha256 of the source cell.

Layout (setup without the gui-agent-research repo)

pip install -U huggingface_hub
python -c "from huggingface_hub import snapshot_download as d; d('yanghs1018/ti_5k', repo_type='dataset', local_dir='ti_5k')"
cd ti_5k/images && for t in part-*.tar; do tar -xf "$t" && rm "$t"; done && cd ../..

Result -- everything lives under one directory, images in a single flat folder:

ti_5k/
  README.md  stats.json
  cells/ti_5k.jsonl         one row per (step, negative)      -- Qwen3.5-9B only
  cells/ti_5k_venus.jsonl   the same rows, UI-Venus-2 format  -- UI-Venus-2-9B only
  images/<uuid>.png          AgentNet frames
  images/<tid>_<t>_<...>.png synthetic image negatives (o^-)      (48176 files in images/)

Every image path inside the cells (images, images_neg, synth.neg_png) is relative to the dataset root, e.g. images/9c5b1bde-....png. Resolve it as os.path.join(<root>, path) (or run from <root>); no other path rewriting is needed. hf_pack.py fetch (below) does the same download/unpack and additionally rewrites the paths to absolute ones for the trainer. Disk: about 30 GB. Unpacking one shard at a time and deleting it (as above) keeps the peak near 30 GB + one shard (~2.5 GB).

import json, os
root = "ti_5k"
for line in open(os.path.join(root, "cells", "ti_5k.jsonl")):
    r = json.loads(line)
    pos_imgs = [os.path.join(root, p) for p in r["images"]]
    neg_imgs = [os.path.join(root, p) for p in r.get("images_neg") or r["images"]]   # text negatives reuse the window

Train

python research/screen-grounding/datagen/scripts/hf_pack.py fetch --repo <user>/ti_5k --out <dir>
bash research/screen-grounding/contrastive-dpo/scripts/train_cell.sh --exp <dir> --cell ti_5k --name <run> --loss pairsum

fetch rewrites the image paths to absolute ones under <dir>/images; the trainer reads cells/ti_5k.ref.jsonl. The commands above are the Qwen3.5-9B path (cells/ti_5k.jsonl). UI-Venus-2-9B (research repo): ref with datagen/scripts/ref.py --backend venus --chains cells/ti_5k_venus.jsonl, train with contrastive.py --backend venus on that ref file. Outside the research repo use training-handoff (--model qwen35 / --model venus), which picks the right cell for you. To get the plain state-line format: refmt.py --cell cells/ti_5k.jsonl --out <x> --fmt plain (list -> plain needs nothing else; plain -> list needs the annotation steps), then recompute the ref with datagen/scripts/ref.py.

Source: AgentNet (xlangai/AgentNet) trajectories, annotated with claude-sonnet-5-5 (v7 whole x diff); image negatives synthesised by image-neg (OmniParser-v2 + EasyOCR elements, retrieval bank, render_neg). Packed 2026-11-05T14:44:28.