DPO-Qwen3-LiteOS-ecb96fbe6

Trajectory-level DPO preference pairs for GUI computer-use agents, built on lite.osworld (train.perturb) and tokenized for Qwen/Qwen3-VL-4B-Instruct.

Generated at cua-lite commit ecb96fbe6.

What a row is

One (chosen, rejected) trajectory pair for the same task, both starting from the same initial environment state. Trajectory-level DPO scores a whole trajectory as the sum of its per-action log-probabilities, each action conditioned on its own protocol-windowed context:

S(tau) = sum_t log pi(a_t | h_t),   h_t = f_protocol(x, o_<=t, a_<t)
L_DPO  = -log sigmoid(beta * [ (S_pol(tau+) - S_ref(tau+)) - (S_pol(tau-) - S_ref(tau-)) ])

Only assistant action tokens are scored; observations/screenshots are conditioning context and must be masked out of the loss. The trainer must SUM response_tokens log-probs across every step of a side to reconstruct S(tau) — averaging per step yields a different (weaker) objective.

Schema

column	type	meaning
task_id	string	shared task; both sides ran the same initial state
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]	chosen steps followed by rejected steps, one per assistant action
n_chosen_steps	int64	split point: steps[:n] chosen, steps[n:] rejected
chosen_metadata, rejected_metadata	string	JSON, source rollout metadata

Each steps entry: {prompt, image_indices, response, response_tokens, reward, status, prompt_tokens}. image_indices addresses 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).

The two sides are concatenated into single steps/processed_images columns rather than four chosen_*/rejected_* columns so the row fits training harnesses that expose only a few per-row slots to a rollout function.

Construction

collect -> annotate (quality gates -> metadata.others.exclude_reason)
        -> pair (task_id match, margin filter, action-hash dedup)
        -> export (render + tokenize with the target model's adapter)

• chosen: any trajectory with a clean exclude_reason and the higher return (GPT-5.5 teacher, or a Qwen3-VL-4B rollout that succeeded)
• rejected: Qwen3-VL-4B-Instruct rollouts only — on-policy negatives for that model
• margin > 0.5 (on this env's 0/1 rewards, "success vs failure")
• at most 4 pairs per task, widest margin first
• trajectories with an identical assistant-action sequence are collapsed before pairing, so no duplicate pair inflates one trajectory's gradient weight

Statistics

pairs	382
distinct tasks	109
distinct chosen / rejected trajectories	239 / 224
exact duplicate pairs	0
cross-model / same-model pairs	334 / 48
chosen steps (mean / median)	10.0 / 8
rejected steps (mean / median)	15.8 / 13
response tokens, chosen / rejected	219,933 / 352,971
images	9,850

Rejected-side quality tags (why each negative is a negative) are preserved in rejected_metadata.others.exclude_reason: footgun:loop, incomplete, footgun:undo_storm, etc. A clean tag means the trajectory simply failed the task checker.

Known characteristics

• Length asymmetry: rejected trajectories carry ~1.6x the response tokens of chosen (median per-pair ratio 1.48; 96 of 382 pairs have a shorter rejected side). Since standard DPO sums rather than length-normalizes, some of the achievable margin is attributable to length. Consider tracking premature-termination rate as an explicit eval.
• No Thought: channel. Both sides render as Action: <text> + <tool_call>. The teacher's inline reasoning exists upstream but is not emitted by this adapter configuration, so it is absent from both sides — symmetric, hence no style shortcut, but DPO cannot teach reasoning here.
• 6 chosen trajectories end on a text turn with no tool call (question-answering tasks where the teacher replied in prose instead of calling answer()).
• train.perturb tasks are perturbed variants of lite.osworld eval tasks. Training on this data and evaluating on that eval split is a leakage path.

Provenance

Teacher rollouts: GPT-5.5. Student rollouts: Qwen/Qwen3-VL-4B-Instruct. Tokenizer/chat template frozen at export: Qwen/Qwen3-VL-4B-Instruct — a set exported under one model family's adapter cannot train another.