Spatial CAPTCHA SFT v2
Trajectories of the reference solver on the spatial CAPTCHA bench, each played in the browser and judged solved by
the page, as ms-swift samples for Qwen3.5 (megatron sft --agent_template qwen3_5). Training script and docs: GitHub
zhaihaotian/spatial-captcha-sft.
| tasks | episodes | |
|---|---|---|
| main tasks | 100 | 100,000 |
| 4D tasks (16-frame scenes, stepped with the page's Next frame / turn buttons) | 25 | 25,000 |
- Reflect: 16 tasks have, among their 1,000 episodes, 300 that make one deliberate mistake and fix it after
looking at the screen: 12 held-drag tasks (the first drag lands off on purpose, then is pulled into place) and 4 step
tasks (
silhouette_build,tower_crane_set,roll_cube_red_up,hinge_endpoint_transfer: one wrong step, then the fix). The mistake turns have"loss": false. - Next frame presses: in
train_reflect.jsonlandval.jsonl(2026-11-09), in the 15 4D tasks stepped with Next frame and forf16_robot_vac's Wait, only the first press of a row is trained; the rest of the row has"loss": false(168,295 calls). A model trained on them answered 4D tasks before stepping through the frames, so fromtrain_300.jsonl/val_300.jsonl(2026-11-10) on, every press is trained. - Layout: every episode is played on its own random layout: the card anywhere on the 1920x1080 screen (at least 12 px from every edge), drawn 1.0 to 1.2 times larger, on one of 10 page colours.
| file | episodes | samples | images |
|---|---|---|---|
train_reflect.jsonl |
base + reflect | 124,625 | 826,611 |
val.jsonl |
3 of each of the 125 tasks, from base (whole episodes) | 375 | 2,485 |
train_300.jsonl |
300 episodes of each of the 125 tasks drawn from train_reflect.jsonl (seed 0; 36,100 base + 1,400 reflect), every Next frame press trained |
37,500 | 248,587 |
val_300.jsonl |
the samples of val.jsonl, every Next frame press trained |
375 | 2,485 |
train_2d_1000.jsonl |
every episode of the 100 main (2D) tasks in train_reflect.jsonl (94,900 base + 4,800 reflect), no 4D task |
99,700 | 497,065 |
val_2d.jsonl |
the 300 samples of val.jsonl from the 100 main tasks |
300 | 1,481 |
train_2d500_4d200.jsonl |
500 episodes of each main (2D) task and 200 of each 4D task (4D episodes are ~2.3x as long, so each task gets about the same number of trained turns), drawn from the every-press-trained export (seed 0; 52,548 base + 2,452 reflect), every Next frame press trained | 55,000 | 315,363 |
Context: the last 20 screenshots
Each turn is chosen on the last 20 screenshots, a window that slides one screenshot a turn, as the evaluation's Qwen
agent sees them (image_max 20, fold_size 1). Every earlier action stays as text; screenshots older than the
window, the first one included, read "This screenshot has been collapsed." (the first user turn then keeps only the
instruction).
No episode here has more than 18 turns, so every episode is one sample with all its screenshots, and all its turns are trained (except the ones above). An episode of more than 20 turns would give one sample for its first 20 turns and one more sample for each later turn, with the turns before it as context (loss false). (v1 cut episodes into 10-turn blocks and always kept the first screenshot.)
Files
| path | what |
|---|---|
train_reflect.jsonl, val.jsonl |
the samples, frames at 1280x720 (images_720p/...jpg) |
*.meta.jsonl |
one line per sample: id, split, task, turns (the trained turns, [from, to)), n_turns |
val_episodes.json |
the validation episodes |
images_720p/<split>/<task>.tar |
the frames, 1280x720 JPEG (q90, 4:4:4); split base, reflect |
seeds_base.json, seeds_reflect.json |
the instances: task, seed, level, knobs |
train_plain.jsonl, images_720p/plain/, seeds_plain.json |
left from an earlier comparison: not used, loss flags not updated |
Unpack the tars at the dataset root (the image paths are relative to it) and train from there:
for f in images_720p/*/*.tar; do tar -xf "$f"; done
megatron sft ... --dataset train_reflect.jsonl --val_dataset val.jsonl --agent_template qwen3_5 --max_length 20480
Lengths
With the Qwen3.5 tokenizer (an image is 880 tokens at 1280x720, the tools turn 1,032): a sample has 7,330 tokens on
average and at most 18,113 (4D tasks 13,593 on average, main tasks 5,777); 6.6 images on average, at most 18. About
0.92 B tokens a pass over a training set. Samples over --max_length are dropped (truncation_strategy delete), so set
it above 18,113.
Sample format
As v1: no system message; the one tool, computer_use (19 actions, coordinates on a 0-999 grid), is in each row's
tools field. The first user turn is the first screenshot and the bench's instruction; each later screenshot is a
tool_response. Consecutive tool_call messages are one assistant turn, ending where the next move needs a new
screenshot (a click; a drag as mouse_move + left_click_drag; a held drag as mouse_move + left_mouse_down + the
first move, then one move a turn, then left_mouse_up). Every episode ends with the click on Verify. Thinking is off.
- Downloads last month
- 3,152