DPO-Qwen3-MobileGym

Trajectory-level DPO preference pairs for mobile GUI agents, rendered and tokenized for Qwen/Qwen3-VL-4B-Instruct on the mobilegym environment.

What this is

336 (chosen, rejected) trajectory pairs over 89 mobilegym tasks. Both sides of a pair share a task_id; the chosen trajectory scored a higher episode_return than the rejected one by more than a 0.5 margin.

This is trajectory-level DPO data, not final-step DPO. Each side is stored as a list of per-action (prompt, response) steps, each rendered with the context the model would actually see at that turn (history-windowed by the agent's protocol — here Qwen3-VL's rolling 4-image window plus a text summary of older turns). The intended objective is

S(tau) = sum_t log pi(a_t | h_t)
L_DPO  = -log sigmoid(beta * [ (S_pol(tau+) - S_ref(tau+)) - (S_pol(tau-) - S_ref(tau-)) ])

so a trainer must sum per-step response log-probs across every step of a side to get that side's trajectory score. Only assistant-action tokens are scored; observations/screenshots are conditioning context.

Schema

One row per pair. The two sides are concatenated into single steps / processed_images columns and split by n_chosen_steps:

column	type	meaning
task_id	string	mobilegym task
margin	float64	pos_return - neg_return
pos_return, neg_return	float64	episode returns of each side
processed_images	large_list[large_binary]	PNG bytes: chosen's images, then rejected's
steps	list[struct]	steps[:n_chosen_steps] = chosen, steps[n_chosen_steps:] = rejected
n_chosen_steps	int64	split point
chosen_metadata, rejected_metadata	string	JSON LiteMetadata of each source trajectory

Each steps struct: {prompt, image_indices, response, response_tokens, reward, status, prompt_tokens}. image_indices address the concatenated processed_images directly (the rejected side's indices are already offset by the chosen side's image count). prompt_tokens is null for image-bearing steps — vision-token expansion must happen at train time with the processor.

pairs.json is the provenance manifest: which source rollout file and row each side came from, plus per-side model_id and episode_return.

Provenance

• Environment: mobilegym
• Chosen: mostly GPT-5.5 teacher rollouts (293 pairs); 43 pairs use a Qwen3-VL-4B trajectory as chosen
• Rejected: mostly Qwen3-VL-4B student rollouts (279 pairs); 57 pairs use a GPT-5.5 trajectory as rejected
• Quality annotation via the mobilegym annotate pass; a trajectory carrying any exclude_reason is never used as chosen, but may be used as rejected
• Rendering/tokenization frozen at export against Qwen/Qwen3-VL-4B-Instruct. The chat template and tokenizer are baked in — this data does not transfer to another model family without re-export.

App coverage (pairs): wechat 37, weather 32, redbook 27, clock 24, crossapp_content 24, reddit 24, crossapp_commerce 23, tencent_meeting 18, alipay 16, bilibili 16, calendar 16, crossapp_life 16, wechat_reading 16, sms 14, notes 13, railway12306 8, x 8, ebay 4.

Rejected-side failure modes: footgun:loop 170, clean-but-lower-return 88, self_reported_failure 31, incomplete 29, plus combinations.

Known issues — read before training

1. ~18% duplicate pairs. 60 of 336 rows are exact duplicates (identical chosen AND rejected action sequences). Cause: the student rollout ran --group-size 3 at temperature 0.0, so several samples of a task produced identical trajectories, and pairing deduplicated by source-row reference rather than by content. Deduplicate by content hash before training, or those pairs get double gradient weight.
2. Chosen trajectories are reused. 179 distinct chosen trajectories back 336 pairs (57 used 3x, 10 used 4x). 79 of 89 tasks hit the 4-pairs-per-task cap.
3. 43 pairs use a student trajectory as chosen. Selection was by reward only. A lucky-but-sloppy student trajectory can therefore act as a positive example; filter on chosen_metadata.others.model_id if you want teacher-only positives.
4. Length asymmetry. Rejected sides carry 1.78x the response tokens of chosen overall (median 1.53x, p90 5.49x, max 16.8x; 95 pairs have a shorter rejected side). With unnormalized summed log-probs this biases toward shorter outputs — watch for premature termination.
5. 3 chosen trajectories have no submit action (they end on a click despite terminated=true and episode_return=1.0).
6. Margin threshold discards partial-credit pairs. mobilegym rewards are fractional (0.0/0.333/0.667/1.0); the 0.5 margin drops every "both partly right, one better" pair. 320 pairs have margin 1.0, 16 have 0.667.
7. Narrow task coverage — 89 distinct tasks.

Verified

Every row passes: image-placeholder count matches image_indices on all steps, all prompts end at the assistant generation header, all responses end with <|im_end|>, image indices are contiguous with no gaps, and the chosen/rejected image ranges do not overlap.