You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ORA teacher trajectories

New to this data? Start with rua-dev docs/onboarding.md §4: how to read the tables, conversations/, prefixes/, reconstruct.py and the image tars; §5 covers reading a sealed run with ORA itself.

Private repository. Every robot episode the two teacher models ran under ORA (the robot agent harness) on the LIBERO-Pro, RoboTwin, RoboCasa and RoboLab simulation benchmarks, successful and failed alike, with every model call: visible text, the full reasoning_content, tool calls, what the tools answered, what the robot did, and enough to rebuild byte for byte the exact request each call saw (reconstruct.py). One directory per teacher, same layout:

Directory Teacher Runs packed Model calls Images Licence of the outputs
deepseek-1017/ DeepSeek-V4.1-Flash-1017 705 32,523 68,706 usable: self-hosted weights under the MIT licence; the outputs may be used
flash-next/ Qwen3.8-Flash-Next 538 26,488 53,947 under review: the weights are under the Qwen Community License 1.0 and whether their outputs may be used or shared is still being reviewed

Flash-Next's outputs are under licence review: until that review closes this repository stays private and the flash-next/ directory is not to be redistributed or trained on outside the team. DeepSeek-1017's outputs are usable (self-hosted weights under MIT). Keep the two apart when you pass data on.

Layout of a teacher directory

File What
runs.parquet / runs.jsonl One row per run of the teacher found under the outputs roots (packed or not, with the reason)
steps.parquet / steps.jsonl.zst One row per model call of a packed run
conversations/<run_id>.jsonl.zst The run's messages, each stored once in the order it first reached the model, with ORA's later forms of it as variants; the first line is the run header with every step's request layout
prefixes/texts.jsonl.zst, prefixes.index.parquet Texts shared across steps and runs by sha256: system prompt, instructions (evidence invariants + embodiment manual), task sentences
prefixes/tools.jsonl.zst Tool definitions (the tools array of a request) by sha256
images/images-NNN.tar, images.index.parquet Every distinct image the model saw, images/<aa>/<sha256>.png, with source camera or tool, first run and step
inventory/, stats.json The inventory scan as written; the teacher's statistics

Top level: README.md (this card), reconstruct.py, manifest.json (sha256 of every file), stats.json, verify-report.json.

Versions: this is the v2 layout (created 2026-10-23). A v1 of flash-next/ was uploaded earlier the same day (381 runs, per-step requests/<run_id>.jsonl.zst instead of conversations and prefixes); v2 replaced it and it remains only in the repository's commit history.

Why conversations, not requests

ORA sends the whole history again at every step, so the archived canonical requests are about 95% repetition. The pack stores each run once: conversations/<run_id>.jsonl.zst holds one entry per distinct message (task, initial observation, each assistant turn with its reasoning_content, each tool result, each new observation; images as sha256: references) in the order the model first saw it. When ORA later showed a message in another form (a tool result externalised to a preview with a <recoverable_artifact> pointer, an observation whose images aged out, a re-serialised history turn), that form is kept as a variant of the same entry. The run header lists, for every step, the request parameters, the tools sha256 and the layout: the request's messages as references into the conversation (i = entry i's first form, [i, v] = its variant v). reconstruct.py turns a layout back into the exact canonical request (json.dumps(..., sort_keys=True, separators=(",", ":")), ORA's canonical_bytes) and checks its sha256 against the archive's; verify-report.json records that check on 50 sampled steps per teacher.

Entry fields: index, kind (system, instructions, task, static_facts, observation, tool_images, assistant, tool_result, checkpoint, other), role, model_call_id and observation_step (the step that produced or first consumed it), first_call_index, in_request (false for an assistant turn no later request showed: the last one, or a truncated answer ORA did not keep), sha256 (of the stored form), tool_call_id, images, message (the stored form; a prefix text appears as {"$prefix_sha256": …}), variants; assistant entries also carry reasoning_content, finish_reason and usage.

runs columns

run_id, teacher_key, root (which outputs root held the run), run_dir (relative to that root), namespace, experiment, config, backend (libero_pro, robotwin, robocasa, robolab), suite + task_id (LIBERO-Pro) or task + task_config (RoboTwin), seed, environment_driver, required_revision, partition, instruction (the task sentence), episode_identity (the held-out identity key rua-dev uses), teacher, model, provider_profile, reasoning_effort, max_output_tokens, context_window_tokens, tool_surface, tool_surface_composition_sha256, context_profile, memory_policy, budgets (JSON), state, official_success (the benchmark's own verdict, never the model's claim), termination, termination_authority, policy_finish_outcome / _category / _reason, decisions (robot actions executed), model_calls, refused_proposals (zero-write refusals), wall_duration_s, started (UTC), date, sealed, sealed_source_commit, sealed_bytes (the sealed inventory's bytes), evaluation_status, learning_eligible, multi_round, excluded_reason, round1a_v6_selected, round1a_v6_samples, in_round1a_v6, packed, forbidden_path, and the pack's own counts: steps, image_parts, unique_images, conversation_entries, conversation_variants, conversation_bytes, disk_bytes (the run directory on disk), input_tokens, output_tokens, reasoning_tokens, cached_input_tokens, tool_errors, actions, refused_computed, finish_length_steps, not_accepted_steps, requests_unavailable.

steps columns

One row per provider exchange (model call), in call order.

Column Meaning
run_id, model_call_id, call_index, observation_step The call; call_index counts from 0 within the run
request_unix_s, latency_s When the request was sent; seconds until the response event
content The assistant's visible text (often empty: the teacher puts its thinking in reasoning_content)
reasoning_content, reasoning_chars The full chain of thought the provider returned, verbatim
finish_reason, accepted, attempts Provider finish reason; whether ORA accepted the message; provider attempts
decision_kind robot_action (a mutation tool was called), lifecycle (finish_episode), read_only_or_reflection
tool_names, n_tool_calls, tool_calls_json The tool calls: [{id, name, arguments}], arguments as the JSON string the model wrote
tool_results_json, n_tool_errors Per call: ok, status, error_code, error_message, result_text (first 4,000 characters; the full text is the conversation's tool_result entry), result_chars, result_ref (the result artifact in the run)
action_id, n_actions, action_tool The robot action this step proposed (at most one per step)
action_execution_status, action_failure_code, action_failure_stage, action_state_changed From the environment receipt: reached, not_reached, executed, planner_failed, plan_rejected, completed (a grasp skill sequence), …
action_refused True when the planner or Supervisor refused the action without writing to the world (no robot decision spent; ORA's refused_proposals)
action_receipt_ref The receipt artifact in the run
last_action_effect_summary, last_action_effect_json The effect of this step's action as the model read it in the next request (last_action_effect)
new_image_sha256s, n_images_in_request Images that appeared for the first time in this run in this step's request; images inline in the request
messages_in_request, n_history_variants, new_conversation_entries, context_checkpointed Request length; history messages shown in a later form (externalised, aged); messages first seen here; whether the request carries an <emergency_context_checkpoint> (ORA compacted the context)
usage_*, usage_json Provider usage: input, output, reasoning, cached input, total tokens
estimated_input_tokens, emergency_compactions, episode_checkpoints From ORA's context projection record for the call
request_sha256, request_bytes, request_path, request_archived, request_available The canonical request artifact (sha256 of its bytes; in the run's native/provider-requests.tar.zst when archived)
tools_sha256, layout_json The request's tool definitions (prefixes/tools) and its message layout over the conversation
request_event_id, response_event_id, response_event_sha256, response_artifact_ref The agent-trace events the row was read from

images.index

sha256, bytes, width, height, format, path (inside the tar), tar, source_kind (observation, tool, historical, other), camera (observation images), tool (tool feedback images), label (the text that preceded the image in the request), first_run_id, first_model_call_id, first_call_index, occurrences (inline image parts over all requests: an image stays in the context for several steps), runs.

Loading

import json, pandas as pd, zstandard
T = "deepseek-1017"                      # or "flash-next"
runs = pd.read_parquet(f"{T}/runs.parquet"); steps = pd.read_parquet(f"{T}/steps.parquet")
packed = runs[runs.packed]
print(packed.groupby(["backend", "official_success"]).size()); print(steps.reasoning_chars.describe())
calls = json.loads(steps.iloc[0].tool_calls_json)
# one run's conversation: header first, then the entries
run_id = packed.run_id.iloc[0]
with open(f"{T}/conversations/{run_id}.jsonl.zst", "rb") as f:
    header, *entries = [json.loads(l) for l in zstandard.ZstdDecompressor().stream_reader(f).read().decode().splitlines()]
for e in entries:
    if e["role"] == "assistant":
        print(e["model_call_id"], e["reasoning_content"][:200])
# the exact request of step 12, rebuilt (needs the images: a tar in <T>/images/ or an extracted directory)
import subprocess; subprocess.run(["python", "reconstruct.py", ".", "--teacher", T, "--run", run_id, "--call", "12", "--out", "req.json"])
from datasets import load_dataset
steps = load_dataset("<org>/<repo>", "deepseek-1017-steps", split="train")   # private repo: pass token=...

DeepSeek-V4.1-Flash-1017 (deepseek-1017/)

Licence of the outputs: usable. self-hosted weights under the MIT licence; the outputs may be used.

Runs recorded 2026-10-17 to 2026-10-23. Provider profiles: fsmbz-deepseek-1017 (4), fsmbz_dsk1017_effort_high_32k_v1 (29), fsmbz_dsk1017_effort_low_32k_v1 (30), fsmbz_dsk1017_effort_max_32k_v1 (29), fsmbz_dsk41_flash_1017_router_tls_low_16k_v1 (440), fsmbz_dsk41_flash_1017_tls_low_16k_v1 (26), orr_deepseek-1017_dsk1017-888a (1), orr_deepseek-1017_dsk1017-888a-low (81), orr_deepseek-1017_dsk1017-888b (1), orr_deepseek-1017_dsk1017-888b-low (64); reasoning effort: high (31), low (645), max (29). Tool surfaces: libero_placement_direct_overhead_v1 (405), robocasa_mobile_direct_v1 (28), robolab_direct_v1 (6), robotwin_toolchain_b2_closed_noservo_v2 (266). Sealed ORA source commits: 001c63f17c17 (3), 01d1e9cef855 (3), 1f1cb68df40f (3), 22a2f54dc68f (3), 35e9dd6b3620 (2), 519ccf9fe230 (145), 52655ee6674a (3), 5ce8f755e5dd (3), 7ebdd7533418 (1), a20594214857 (2), b44685a360bd (130), b4f7bb79fc54 (88), bc274ad2a3ad (1), c295569ef27a (302), cc3ebe111972 (3), d441d1172e2b (1), e853eec9cc79 (3), e93a9c2fba99 (6), f7ee86ea14a5 (3).

Packed runs (sealed, valid evaluation, single round):

Root Namespace Backend Runs Official successes Success rate Model calls Robot decisions
b2-outputs ora-backends robocasa 28 12 43% 753 354
b2-outputs ora-backends robolab 6 6 100% 128 62
b2-outputs ora-pi-spike libero_pro 1 1 100% 12 7
b2-outputs ora-pi-spike robotwin 1 1 100% 19 7
b2-outputs rpent-open-models libero_pro 204 115 56% 11,929 7,343
b2-outputs rpent-open-models robotwin 98 61 62% 4,484 2,008
ora-outputs ora-native-golden-1021 libero_pro 1 1 100% 27 11
ora-outputs ora-native-golden-1021 robotwin 1 1 100% 33 10
ora-outputs ora-open-models libero_pro 80 65 81% 3,253 1,988
ora-outputs ora-open-models robotwin 138 65 47% 6,463 2,820
ora-outputs ora-structure-smoke robotwin 1 1 100% 32 15
ora-outputs ora-structure-trial robotwin 1 1 100% 9 2
ora-outputs rua-heldout libero_pro 119 97 82% 4,470 2,698
ora-outputs rua-heldout robotwin 26 20 77% 911 409

All runs of this teacher found under the roots, by date window (the team's inventory of 2026-10-20 is the reference; runs started that day may have finished after it was taken):

Root Namespace Backend Window Runs Official successes
b2-outputs ora-backends robocasa before 2026-10-20 32 12
b2-outputs ora-backends robolab before 2026-10-20 6 6
b2-outputs ora-pi-spike libero_pro before 2026-10-20 2 1
b2-outputs ora-pi-spike robotwin before 2026-10-20 2 1
b2-outputs rpent-open-models libero_pro before 2026-10-20 154 52
b2-outputs rpent-open-models libero_pro on 2026-10-20 66 63
b2-outputs rpent-open-models robotwin before 2026-10-20 63 30
b2-outputs rpent-open-models robotwin on 2026-10-20 38 31
ora-outputs ora-native-golden-1021 libero_pro after 2026-10-20 1 1
ora-outputs ora-native-golden-1021 robotwin after 2026-10-20 1 1
ora-outputs ora-open-models libero_pro after 2026-10-20 80 65
ora-outputs ora-open-models robotwin after 2026-10-20 140 65
ora-outputs ora-structure-smoke robotwin after 2026-10-20 1 1
ora-outputs ora-structure-trial robotwin after 2026-10-20 2 1
ora-outputs rua-heldout libero_pro after 2026-10-20 119 97
ora-outputs rua-heldout robotwin after 2026-10-20 38 20

Not packed (40 runs; each is listed in runs.parquet with packed = false, and in stats.json with state and model-call count):

Reason Runs
evaluation not valid 3
exported text names a mount path 1
no result.json (run not finished) 13
not sealed 23

Statistics:

Steps (model calls) 32,523
Steps with a tool call 32,297
Robot actions proposed / refused without a write 18,956 / 1,222
Steps with a tool error 558 (588 errors)
Steps with finish_reason = length 177
Steps whose request carries a context checkpoint 0
Steps without a verifiable canonical request 0
Reasoning length (characters) p50 / p90 / max 4,593 / 21,007 / 115,626
Output tokens per call p50 / p90 / max 1,460 / 6,194 / 32,768
Input tokens per call p50 / p90 / max 175,046 / 350,442 / 646,129
Latency per call (s) p50 / p90 15.811 / 77.582
Steps per run p50 / p90 / max 39 / 93 / 200
Messages per request p50 / p90 / max 83 / 217 / 549
Total tokens: input / output / reasoning 6,168,514,495 / 82,778,085 / 78,150,338
Conversation entries / variants 110,681 / 23,902
Prefix texts / tool-definition versions 221 / 14
Distinct images 68,706 (6.4 GiB)
In round1a-v6's final rows 164 runs, 3,990 samples (6 selected but dropped at merge)

Size: sealed inventory 329.0 GiB, run directories on disk 160.4 GiB, this directory 7.0 GiB (images 6.4 GiB, conversations 0.1 GiB): 47.3× smaller than the sealed inventory, 23.1× smaller than the run directories.

Decision kinds: lifecycle 718, read_only_or_reflection 12,693, robot_action 19,112. Finish reasons: length 177, stop 49, tool_calls 32,297. Action execution status ((not reported): LIBERO-Pro gripper receipts carry none): (not reported) 2,690, aborted 105, completed 297, executed 1,253, not_reached 3,546, plan_rejected 216, planner_failed 924, reached 9,919, terminated 6. Tool calls by name: move_pose 9,719, vision_ground 4,175, move_to 4,051, set_gripper 3,680, geometry_query 2,277, compute 1,899, grasp_plan 1,754, mark_point 1,371, check_task 1,318, vision_segment 1,229, finish_episode 718, grasp_execute 489, align_plan 458, recover_pose 407, observe 282, rotate_wrist 265, rotate_pitch 224, move_base 143, wait_ticks 134, report_issue 16, artifact_read 2. Images by source: observation 47,854, tool 20,852; by camera: agentview 12,797, front_camera 5,389, head_camera 5,685, left_camera 5,567, right_camera 5,507, wrist 12,797; by tool: grasp_plan 3,239, mark_point 1,360, observe 112, vision_ground 12,981, vision_segment 3,272.

Qwen3.8-Flash-Next (flash-next/)

Licence of the outputs: under review. the weights are under the Qwen Community License 1.0 and whether their outputs may be used or shared is still being reviewed.

Runs recorded 2026-10-17 to 2026-10-22. Provider profiles: fsmbz-flashnext (2), qwen38_flash_next_fsmbz_sglang_thinking_16k_v1 (11), qwen38_flash_next_router_tls_thinking_16k_v1 (306), qwen38_flash_next_share_tls_thinking_16k_v1 (157), qwen38_flash_next_sim8_sglang_thinking_16k_v1 (62); reasoning effort: provider_default (538). Tool surfaces: libero_placement_direct_overhead_v1 (318), robotwin_toolchain_b2_closed_noservo_v2 (220). Sealed ORA source commits: 81de90c98704 (2), b44685a360bd (157), c295569ef27a (379).

Packed runs (sealed, valid evaluation, single round):

Root Namespace Backend Runs Official successes Success rate Model calls Robot decisions
b2-outputs ora-pi-spike libero_pro 1 1 100% 33 19
b2-outputs ora-pi-spike robotwin 1 1 100% 13 5
b2-outputs rpent-open-models libero_pro 237 195 82% 12,003 6,572
b2-outputs rpent-open-models robotwin 142 121 85% 6,140 2,784
ora-outputs ora-open-models libero_pro 80 70 88% 3,929 2,116
ora-outputs ora-open-models robotwin 77 50 65% 4,370 2,107

All runs of this teacher found under the roots, by date window (the team's inventory of 2026-10-20 is the reference; runs started that day may have finished after it was taken):

Root Namespace Backend Window Runs Official successes
b2-outputs ora-pi-spike libero_pro on 2026-10-20 1 1
b2-outputs ora-pi-spike robotwin on 2026-10-20 1 1
b2-outputs rpent-open-models libero_pro before 2026-10-20 252 176
b2-outputs rpent-open-models libero_pro on 2026-10-20 34 19
b2-outputs rpent-open-models robotwin before 2026-10-20 160 121
b2-outputs rpent-open-models robotwin on 2026-10-20 7 1
ora-outputs ora-open-models libero_pro after 2026-10-20 83 70
ora-outputs ora-open-models robotwin after 2026-10-20 80 50

Not packed (80 runs; each is listed in runs.parquet with packed = false, and in stats.json with state and model-call count):

Reason Runs
no result.json (run not finished) 11
not sealed 69

Statistics:

Steps (model calls) 26,488
Steps with a tool call 26,326
Robot actions proposed / refused without a write 14,577 / 974
Steps with a tool error 915 (926 errors)
Steps with finish_reason = length 155
Steps whose request carries a context checkpoint 12,932
Steps without a verifiable canonical request 0
Reasoning length (characters) p50 / p90 / max 4,980 / 19,289 / 62,589
Output tokens per call p50 / p90 / max 1,759 / 6,426 / 16,384
Input tokens per call p50 / p90 / max 124,515 / 180,123 / 233,822
Latency per call (s) p50 / p90 23.724 / 87.757
Steps per run p50 / p90 / max 33 / 122 / 186
Messages per request p50 / p90 / max 48 / 89 / 183
Total tokens: input / output / reasoning 3,267,063,272 / 71,763,910 / 0
Conversation entries / variants 89,010 / 16,602
Prefix texts / tool-definition versions 208 / 2
Distinct images 53,947 (4.8 GiB)
In round1a-v6's final rows 277 runs, 8,195 samples (6 selected but dropped at merge)

Size: sealed inventory 237.0 GiB, run directories on disk 114.8 GiB, this directory 5.2 GiB (images 4.8 GiB, conversations 0.1 GiB): 45.1× smaller than the sealed inventory, 21.9× smaller than the run directories.

Decision kinds: lifecycle 536, read_only_or_reflection 11,250, robot_action 14,702. Finish reasons: invalid_tool_input 2, length 155, stop 5, tool_calls 26,326. Action execution status ((not reported): LIBERO-Pro gripper receipts carry none): (not reported) 1,394, aborted 45, completed 155, executed 1,022, execution_fault 3, not_reached 2,312, plan_rejected 112, planner_failed 795, reached 8,735, terminated 4. Tool calls by name: move_to 5,836, move_pose 5,650, mark_point 2,843, compute 2,677, vision_ground 2,488, set_gripper 2,277, geometry_query 1,607, vision_segment 1,345, check_task 733, grasp_plan 707, finish_episode 536, rotate_wrist 313, grasp_execute 246, align_plan 241, rotate_pitch 173, recover_pose 162, observe 67, wait_ticks 45, artifact_read 15, report_issue 6, interaction 1. Images by source: observation 38,333, tool 15,614; by camera: agentview 8,933, front_camera 5,030, head_camera 5,186, left_camera 5,118, right_camera 5,098, wrist 8,933; by tool: grasp_plan 1,378, mark_point 2,819, observe 35, vision_ground 7,817, vision_segment 3,600.

Known issues and reading notes

  • reasoning_content is the provider's reasoning stream, not something ORA edited. usage_reasoning_tokens is 0 where the server did not report it (the SGLang-served Flash-Next profiles); use reasoning_chars or usage_output_tokens.
  • finish_reason = length: the output budget ran out; the message usually has no tool call and ORA asked again on the same observation (provider_output_truncated).
  • Refused actions (action_refused): the planner or Supervisor refused the proposed motion before any write (planner_failed, plan_rejected). The world did not change and no robot decision was spent; the model read the refusal in the next request's last_action_effect.
  • Tool errors (error_code in tool_results_json): mostly invalid_arguments (schema or budget violations in compute, move_to, …). The model saw the same error text.
  • Context checkpoints: long episodes were compacted by ORA's compactor (the same model summarising its own history); the request then starts with an <emergency_context_checkpoint> message instead of the full early history. The steps before the checkpoint are still in steps and in the conversation.
  • Images: the model saw these exact bytes (PNG; 256×256 or 320×240 observations, up to 1024×800 tool previews). ORA shows recent images only; older ones become <recoverable_artifact> text pointers in the request.
  • Requests unavailable: a step without a verifiable canonical request (request_available = false) keeps its response row and its assistant entry; its layout is empty.
  • Multi-round, unsealed and unfinished runs are listed in runs but have no steps.
  • Not included: runs of the official DeepSeek API model deepseek-flash (s1s2-pilot) and of DeepSeek-V4.1-Flash without the 1017 tag: they are other model versions. A run whose exported text names a cluster host, mount path or credential is excluded whole (excluded_reason), like a privileged-field hit: this removed the interactive ora-play session, whose coding-agent system prompt lists cluster paths.

Relation to round1a-v6

rua-dev's SFT round 1a (v6, built 2026-10-22) selected official successes on the default tool surfaces, at most 2 runs per (episode identity, teacher) and 4 per identity, and dropped RoboTwin tasks whose checker is loose (place_bread_skillet, shake_bottle, shake_bottle_horizontally). in_round1a_v6 marks the packed runs whose samples are in round1a-v6's final rows (Flash-Next's under the licence-review split that is not trained on while the licence is open; DeepSeek-1017's under usable); round1a_v6_selected without in_round1a_v6 marks runs selected but dropped at merge. Everything else here (failed episodes, non-default surfaces, capped duplicates, runs after 2026-10-22) never entered round 1a.

Privacy and safety

Every exported text (requests, tool results, responses) passed ORA's privileged-field scan (_forbidden_path: no object_gt, segmentation, contact, success_predicate, actor_state, expert_program, task_source fields) and a scan for cluster hosts, mount paths, bearer tokens, key material and API-key literals; a run with a hit is excluded whole and named in runs.excluded_reason (verify-report.json repeats the scan on the pack). Run directories are given relative to the team's outputs roots (root, run_dir).

Provenance

Produced by sft/tools/teacher_pack.py in rua-dev, which reads runs through ORA's own evidence helpers (provider-exchange pairing, the request archive, the privileged-field scan). ORA line at read time: backends; ORA.lock commit 55a7353bf35a. Created 2026-10-23T05:42:29Z.

Downloads last month
393