Raw preference candidates and web trajectories with mixed per-source benchmark licenses. Access is reviewed manually. Please describe intended use and preserve evaluation exclusions.

Web Trajectories Preference Raw

A raw inventory for later filtering, not a ready-to-train preference corpus. It preserves 8,851 successful and 5,960 failure-repository sessions, with 618,539 original LLM-call records.

There are 423 tasks in both repositories, yielding 81,850 task-level session combinations. These are not all distinct preference judgments. task_pairs identifies complete failures, incomplete runs and conflicting success flags separately.

exact_prompt_pairs contains 36,712 content-deduplicated response pairs whose full original messages, tools hash, role and purpose match exactly. shared_prompts stores each matched prompt once. Join on prompt_id; resolve tool definitions using tools_hash in the original tool tables.

Preference levels

Config	Meaning	Limitation
task_pairs	Same benchmark source/task ID, success/failure session references	Different observations/histories; not automatically a same-prompt DPO pair
exact_prompt_pairs	Byte-identical full request messages and tool schema hash, different meaningful responses	Episode outcome does not prove each local action preference; review required
executed_recovery_pairs	4 same-context corrections actually executed; 3 state changes and 1 fresh observation	Local recovery only, 2 tasks, not independent human gold
counterfactual_pairs	4 previously reviewed correction proposals	Chosen corrections were not executed
derived_examples / derived_membership	Existing SFT experiments, content deduplicated with repetition provenance	Positive SFT targets alone do not form preference judgments
raw_*	Complete original session, turn and tool parquet snapshots	Includes every original flag, incomplete run and historical tool version

All candidate tables are marked training_ready=false. local_execution_status_preference is a mechanical signal, not semantic or human validation. Random provider tool-call IDs are removed for response-content deduplication; observations, transient target IDs, actual tool arguments and full original prompts are not rewritten. Original responses remain available in raw tables.

Evaluation boundaries

partition is one of train, development, protected_evaluation, unassigned. Original SFT test tasks, original84 and supplemental24 are protected. Existing dev8 and native-validation tasks are development. Only partition=train allows later training selection, and that still requires preference review. Keep task/session boundaries across any new training/validation split. Unassigned tasks are excluded until a deliberate split policy is defined.

No held-out data was used to train a model or choose a new model here. This release is an archival/raw inventory. Do not load the entire raw split straight into a trainer. No training was run or scheduled.

Schema and usage

from datasets import load_dataset
pairs = load_dataset("lukakuma/web-trajectories-preference-raw", "task_pairs", split="raw", token=True)
candidates = pairs.filter(lambda r: r["partition"] == "train" and r["outcome_class"] == "judged_completed_failure" and r["same_goal"])

For DPO, first select exact-input and semantically supported candidates, join their prompt and tools, and convert chosen/rejected assistant outputs with the model's verified chat template. Tool/environment observation tokens are context and must not be treated as assistant action targets. Trajectory-level preference optimization requires explicit trajectory/probability masking support; task-ID matching alone does not provide that.

statistics.json gives exhaustive counts. sources/file-inventory.json records each pinned upstream parquet SHA256 and row count. sources/revisions.json records upstream commits. SHA256SUMS.json covers release files. Per-source benchmark licenses are preserved in LICENSE; no broader license grant is implied.

Sources

Repeated evidence and historical exposure

exact_pair_evidence preserves all 40,140 supporting session/turn occurrences behind content-deduplicated pairs. exact_pair_quality aggregates whether any occurrence has a complete failure, a local execution-status contrast or known matching software/judge. Main pair rows contain one representative occurrence; consult evidence and quality before filtering.

sft_membership records 372,292 historical training uses in v7a/b/c/d, including v7d loss masking. derived_membership separately preserves later experimental repeats. These are provenance/exposure counts, not extra preference judgments.

Recent local collections

local_collection_sessions and local_collection_turns preserve 32 recent sessions and 630 calls from student collection, teacher collection, native validation and live guided recovery. Unknown outcomes remain unknown; guidance-enabled runs are marked. executed_recovery_evidence contains the4 actually measured correction receipts and their quoted input evidence.

Response comparison canonicalization keeps text and tool-call content and removes provider IDs/state metadata; all original blocks remain in raw call tables. Most current exact-input pairs concern goal decomposition, not locally proven action choices.

Current-contract and curation annotations

Exact-input candidates expose chosen/rejected structured tool names and whether those names exist in frozen actor931ff5cc0's56-tool surface. This is a name-only compatibility check, not argument or semantic validation. curation_flags links8 historical examples excluded from the clean weekend SFT:6 edited-but-unexecuted reports and2 unsupported screenshot targets. They remain archived, not silently deleted.

Within-trajectory recovery transition candidates (v2 extension)

recovery_transition_candidates adds 21,837 raw candidates from 468 TRAIN tasks and 8,116 sessions after scanning all TRAIN sessions in both upstream repositories. Each candidate's current request contains the latest external tool receipt with explicit ok=false; the changed chosen action has its own later, strictly matched ok=true tool receipts. 20,710 candidates occur after the first five LLM calls. All source tool contracts, model/version/judge identifiers, original current messages, receipt contents and IDs are retained.

These are not independently executed same-input chosen/rejected pairs. The chosen action was executed under the stored current request. The rejected action copies the specific earlier failed request into that current request as a counterfactual repeat. It was not executed under the current request. Therefore same_prompt_verified=false, rejected_executed_in_pair_context=false, requires_review=true, and training_ready=false. Do not merge this config with exact-input or execution-verified local-recovery evidence without preserving these distinctions. Repeating an action can sometimes be valid.

There are 8,347 candidates with a recorded meaningful-state-change flag, which is a mechanical signal and not a semantic judgment or proof of task completion. 11,070 chosen source turns were absent from all historical v7a/b/c/d active-loss memberships; 21,668 current-message/response combinations were absent from the archived later experimental derived examples (the latter comparison ignores tools and provider response IDs). These novelty counts are not new gold labels. Teacher model origins remain explicit; the pool is not on-policy Minion failure collection.

The source episode may itself have succeeded or failed; local normal tool status does not upgrade its episode verdict. All candidates remain unapproved and no training was launched. recovery_transition_manifest.json describes extraction, receipt coverage, exclusions, source hashes and exposure counts. The existing exact-input candidates and four executed local corrections remain unchanged. Per-source licenses and protected-evaluation exclusions continue to apply.

Minion-native executed corrections (v3 extension)

minion_native_executed_recovery adds 4 unapproved correction pairs on 3 TRAIN tasks, collected during 10 READ-only task episodes with Minion 1.2B and all 56 tools. There were 28 bounded teacher reviews; 4 proposals were forwarded, executed with distinct receipts, and passed a quoted-evidence and mechanical local-effect check. Three inputs precede any intervention on their task; the fourth follows an earlier intervention. These episodes are teacher assisted and are not unassisted success-rate measurements.

The rejected response is the actual student-generated response under the stored actor input, not a relocated historical action. The chosen action was proposed by Qwen3.6-27B with a separate review prompt, then forwarded and executed under that original actor input. same_prompt_verified=true refers to this common actor context, not identical student/teacher generation prompts. teacher_prompt_same_as_actor_prompt=false, rejected_executed_in_actor_context=false, training_ready=false, and independent_human_gold=false. The rejected action has an earlier measured failure in the history but was not re-executed after interception.

Full tool schemas come from the captured actual API request. Actor messages match the logged request exactly; the frozen actor logs only tool type/name/description[:200], so restoration requires exact equality with that known projection. No schema is guessed from another run. Current assistant targets, real correction receipts, own execution IDs and source hashes are retained. All 4 pairs render without truncation with the starting Minion tokenizer (maximum arm length 34,269 tokens).

Local state change is not sustained recovery or task success. Three corrections have a complete five-external-call follow-up window; all three encounter an exact original failed action failing again. The fourth window is censored. Original verdicts remain unchanged, and all rows require later selection/review. Earlier configs and source licenses remain preserved. This collection does not establish any model improvement and launched no training.

Native continuation and visual raw evidence (v4 extension)

This extension preserves 6 further executed corrections with measured prior failures, separately from 3 weak follow-up proposals whose original student responses were generated but not executed after interception. native_recovery_pairs_v3_v4 and native_weak_followups retain original actor input, full 56-tool contracts, the student response, forwarded teacher action, own execution result, quotes and evidence tier. The teacher used a separate review prompt. Actor-context alignment is not identical student/teacher generation prompting. Local effects do not establish task success. All rows remain training_ready=false.

multimodal_image_catalog contains 120 content-addressed original PNGs, preserving 276 local frame-file occurrences from 20 TRAIN episodes. multimodal_observations links 170 latest-observation/action candidates across 10 tasks to saved screenshots using exact fresh runtime IDs and own execution receipts. No intervening external observation is skipped, and images are never matched by basename or approximate timestamp. Referenced image files are under assets/native_frames/; decode the PNG associated with image_path and verify image_sha256.

The recorded actor input contained no images. These are later-attached visual views of DOM trajectories, not evidence that a visual policy generated the actions. Recorded inner viewport dimensions, PNG dimensions and uniform content scale validate capture geometry for all 170 rows. Five differ from the configured viewport; the actual browser/capture metadata agrees. No image was resized to force alignment. visual_grounding_candidates contains 1,405 unapproved DOM-supervised target candidates on 54 unique images, restricted to frame 0, unique role/label, actionable hints and fully on-screen bounding boxes. Pixel/CSS transforms do not prove visibility, lack of occlusion, semantic correctness or valid visual preferences. visual_visibility_verified=false, semantic_action_correctness_verified=false, and training_ready=false. Split future training and evaluation by task/domain, not target rows from the same image.

Existing configs, raw labels and manual access gating are preserved. This extension makes no new blanket license claim over depicted website content; originating URLs and task/source provenance are retained under the repository's existing per-source policy. No training or browser-success improvement is claimed. native_visual_extension_manifest.json records source/output hashes and configuration counts.

Downloads last month: 5,911

Total file size:

658 MB