RegBench benchmarks LLM agents that register a post-event SAR image (moving) onto an optical image of the same scene (reference): whole scenes at original resolution (0.30–0.38 m per pixel) from SpaceNet-9 and the BRIGHT disaster collection. Each rollout runs in an offline sandbox with a shell, a persistent Python kernel, the regtool registration toolbox (classical and deep matchers such as MINIMA, RoMa and LoFTR, robust estimation, warping), documentation search, a QGIS desktop it can drive by computer use, and the VISTA visual tools (inspect, read_pixels, history and a notebook, read_working / write_working). The agent must submit a transform, the registered raster, its own verification and provenance. A host-side evaluator scores the transform against hidden manual tie points.

This dataset holds 60 rollouts: 4 cohorts × 15 scene tasks, one rollout per scene and cohort, with the full tool-call sequence, a provider-neutral transcript of what each model saw, the images it was shown, its small output files and the scores.

Cohorts

Cohort	Model	Harness	Commit(s)	Experiments	Caps
`opus55_vista_v0`	anthropic/claude-opus-5.5	v0: the VISTA harness of the published leaderboard (branch claude/vista-reg)	30f4a3c, 5cae58e, f53fa23	vista_20261004, vista_bright_v3_20261006, vista_scenes_20261005, vista_test_scenes_20261005	$18.0, 250 calls, 180.0 min, 32000 output tokens
`gpt61sol_vista_v0`	openai/gpt-6.1-sol	v0: the VISTA harness of the published leaderboard (branch claude/vista-reg)	5cae58e, bb56848	vista_gptsol_20261005, vista_gptsol_bright_v3_20261006, vista_gptsol_test_20261005	$18.0, 250 calls, 180.0 min, 32000 output tokens
`opus55_vista_iter01b`	anthropic/claude-opus-5.5	iter01b: v0 plus `regtool field` (robust local-median displacement field from match files, with cross-validation and an uncertainty raster) and a tool-notes paragraph on how accuracy is judged (branch claude/harness-iter, commit 3ca82d9)	3ca82d9	hi01b_opus_20261008	$18.0, 250 calls, 180.0 min, 32000 output tokens
`gpt61sol_vista_iter01b`	openai/gpt-6.1-sol	iter01b: v0 plus `regtool field` (robust local-median displacement field from match files, with cross-validation and an uncertainty raster) and a tool-notes paragraph on how accuracy is judged (branch claude/harness-iter, commit 3ca82d9)	3ca82d9	hi01b_gpt_20261008	$18.0, 250 calls, 180.0 min, 32000 output tokens

All cohorts ran the native-aligned harness (Anthropic Messages for Opus, OpenAI Responses for GPT) with the full tool profile, VISTA on and reasoning effort high.

Metrics

Every scene has manual ground-truth tie points (2182 over the 15 scenes). For each tie point the evaluator maps the reference point through the submitted transform into the SAR image and measures the distance to the manual SAR tie point, in original reference-image pixels (SpaceNet-9 scorer, error E_k). Where the transform is undefined at a tie point (an invalid registration), that point keeps its georeferencing-only error.

MeanErr = mean(E_k), in px.
S@5px / S@10px = share of tie points with E_k ≤ 5 / 10 px.
RMSE = sqrt(mean(E_k²)), in px, never clipped.

Pooled numbers concatenate the tie points of all scenes, so scenes with more tie points weigh more. The manual tie points themselves disagree by about 5–24 px per scene (leave-one-out floor), so gaps of a few pixels are within ground-truth noise.

Results (pooled over all tie points)

Cohort	Scenes	MeanErr (px) ↓	S@5px ↑	S@10px ↑	RMSE (px) ↓	Invalid	USD
opus55_vista_v0	15	17.1	23.3%	50.8%	33.0	0	66.94
gpt61sol_vista_v0	15	19.5	20.5%	44.8%	35.3	1	23.39
opus55_vista_iter01b	15	11.9	32.1%	61.5%	21.6	0	81.41
gpt61sol_vista_iter01b	15	10.6	38.0%	67.1%	19.0	1	34.39

Without the 4 scenes used while developing the iter01b harness (t_dev_0001, t_test_0006, t_test_0047, t_test_0052):

Cohort	Scenes	MeanErr (px) ↓	S@5px ↑	S@10px ↑	RMSE (px) ↓	Invalid	USD
opus55_vista_v0	11	16.8	19.7%	45.8%	27.7	0	49.93
gpt61sol_vista_v0	11	18.1	18.6%	41.7%	28.2	1	18.31
opus55_vista_iter01b	11	14.1	24.0%	52.7%	25.2	0	60.62
gpt61sol_vista_iter01b	11	13.2	28.3%	56.1%	22.9	1	26.90

Per scene, MeanErr (px):

Task	Dataset	Scene	Tie points	opus55_vista_v0	gpt61sol_vista_v0	opus55_vista_iter01b	gpt61sol_vista_iter01b
t_dev_0001	BRIGHT	bata_explosion/scene_1	161	7.2	7.5	6.1	5.5
t_dev_0006	BRIGHT	jamaica_cyclone/scene_2	68	20.1	28.4	22.6	24.6
t_dev_0011	BRIGHT	noto_earthquake/scene_1	64	22.6	12.2	9.8	11.4
t_dev_0016	SN9	02_train_02	104	17.1	32.7 (invalid)	15.6	23.7 (invalid)
t_test_0001	BRIGHT	california_wildfire/scene_1	221	11.7	11.6	12.0	10.3
t_test_0006	BRIGHT	congo_volcano/scene_1	228	9.5	8.9	5.9	5.8
t_test_0011	BRIGHT	hawaii_wildfire/scene_1	75	25.5	22.5	16.8	11.2
t_test_0016	BRIGHT	jamaica_cyclone/scene_1	28	13.9	16.9	16.5	16.8
t_test_0017	BRIGHT	spain_volcano/scene_1	308	15.9	21.2	15.0	11.1
t_test_0027	BRIGHT	turkey_earthquake/scene_1	111	7.4	11.0	8.0	8.3
t_test_0032	BRIGHT	turkey_earthquake/scene_2	124	16.8	14.0	8.1	7.2
t_test_0037	SN9	02_train_01	151	21.1	15.6	16.4	15.2
t_test_0042	SN9	03_train_01	160	20.9	18.2	17.2	16.6
t_test_0047	BRIGHT	morocco_earthquake/scene_1	132	9.1	37.0	9.3	6.5
t_test_0052	BRIGHT	morocco_earthquake/scene_2	247	36.3	35.9	10.4	5.5

Files

data/runs.parquet (config runs): one row per rollout.
data/steps.parquet (config steps): one row per tool call, in order.
data/conversations.jsonl (config conversations): one transcript per rollout.
images/<run_id>/<sha256>.png: every image a tool returned to the model.
artifacts/<run_id>/: run.json, cost.json, provenance.auto.json (harness-observed provenance), eval_summary.json (whitelisted evaluator fields), WORKING.md (the agent's VISTA notebook) and output/, the agent's small files from its output folder (transform.json, verification.json, provenance.json, scripts, logs, match files ≤ 2 MB). Rasters and larger files are not included, so a displacement_field transform names a field raster that is not here.

from datasets import load_dataset
runs = load_dataset("daidaidaidai1dai/regreg", "runs", split="train")
steps = load_dataset("daidaidaidai1dai/regreg", "steps", split="train")
conversations = load_dataset("daidaidaidai1dai/regreg", "conversations", split="train")

`runs`

Column	Type	Description
`run_id`	`string`	`<cohort>__<task_id>`
`cohort`	`string`	model × harness version
`cohort_description`	`string`	what the harness version is
`model`	`string`	requested model (OpenRouter id)
`served_model`	`string`	snapshot reported by the provider
`provider`	`string`	inference provider the run was pinned to
`experiment`	`string`	RegBench experiment of the rollout
`harness_commit`	`string`	git commit the harness ran from
`harness_dirty`	`bool`	uncommitted harness changes at launch
`source_run`	`string`	rollout folder on the organiser's storage
`task_id`	`string`	scene task
`split`	`string`	dev or test
`scene`	`string`	dataset scene name
`dataset`	`string`	SN9 (SpaceNet-9) or BRIGHT
`reasoning_effort`	`string`	requested effort
`transport`	`string`	messages (Anthropic Messages), responses (OpenAI Responses) or chat
`dialect`	`string`	tool dialect of the native-aligned harness
`caps`	`string`	per-rollout caps (JSON)
`eval_report`	`string`	evaluator report the metrics come from
`mean_err_px`	`double`	MeanErr: mean tie-point error (px)
`median_err_px`	`double`	median tie-point error (px)
`p90_err_px`	`double`	90th percentile tie-point error (px)
`s_at_5px`	`double`	share of tie points with error <= 5 px
`s_at_10px`	`double`	share of tie points with error <= 10 px
`rmse_px`	`double`	sqrt(mean(error²)) (px, unclipped)
`n_points`	`int64`	GT tie points of the scene
`ok_registration`	`bool`	transform defined at every tie point (false: those points carry the georeferencing error)
`sn9_rs_px`	`double`	evaluator's SN9 registration score (mean error, px)
`sn9_err_px`	`list<element: double>`	per-tie-point error (px) in GT point order; no coordinates
`usd`	`double`	API cost (USD)
`input_tokens`	`int64`	prompt tokens, all calls
`output_tokens`	`int64`	completion tokens, all calls
`cached_tokens`	`int64`	prompt tokens read from the provider cache
`wall_s`	`double`	episode wall time (s)
`n_tool_calls`	`int64`	tool calls (= rows in steps)
`n_turns`	`int64`	model responses
`stop_reason`	`string`	why the episode ended
`submitted`	`bool`	deliverables validated by submit
`transform_type`	`string`	family of the submitted transform
`agent_accept`	`bool`	agent's own verdict (verification.json)
`agent_confidence`	`double`	agent's stated confidence
`agent_estimated_median_error_px`	`double`	agent's own estimate of its median error (px)
`n_images`	`int64`	distinct images returned by tools
`n_images_elided`	`int64`	images later dropped from the context

`steps`

Turn-level columns (assistant_text, reasoning, usage_*) repeat on every call of the same turn; deduplicate on (run_id, turn) before summing.

Column	Type	Description
`run_id`	`string`	rollout
`step`	`int64`	1-based tool-call index
`turn`	`int64`	0-based model turn that issued the call (null: harness-forced)
`t_offset_s`	`double`	seconds from episode start to the call
`assistant_text`	`string`	visible text of the issuing turn
`reasoning`	`string`	reasoning text of the issuing turn (Opus: summarized thinking; GPT: reasoning summary, often empty)
`tool`	`string`	tool name
`arguments`	`string`	arguments (JSON)
`result_text`	`string`	full tool result text, incl. the harness budget line
`error`	`bool`	tool reported an error
`wall_s`	`double`	tool execution time (s)
`forced`	`bool`	call made by the harness, not the model
`image_paths`	`list<element: string>`	images the tool returned
`usage_prompt_tokens`	`int64`	prompt tokens of the issuing turn
`usage_completion_tokens`	`int64`	completion tokens of the issuing turn
`usage_cached_tokens`	`int64`	cached prompt tokens
`usage_reasoning_tokens`	`int64`	reasoning/thinking tokens of the issuing turn
`usage_usd`	`double`	cost of the issuing turn (USD)

`conversations`

One JSON object per rollout: run_id, cohort, model, task_id, generation (request settings such as max_tokens, thinking / reasoning effort, as JSON), tools (name, description and JSON-schema parameters as a JSON string, exactly as declared to the model), messages and harness_events. It is rebuilt from the request payloads the harness sent, so it is what the model saw, in order:

system, then user messages: the task text and the harness instructions (VISTA paragraph).
assistant: turn (0-based model call), content (text parts), reasoning (Opus: the provider's summarized thinking; GPT: the reasoning summary, often empty), tool_calls (id, name, arguments as a JSON string).
tool: tool_call_id, name, is_error, content = text part (the tool's text plus the harness budget line) and image parts {"type": "image", "path": "images/<run_id>/<file>"}.
user messages after the first assistant turn have harness_injected: true (budget warnings, "one tool call left", "continue using the tools"). This export has 0 such messages.
elided_from_turn: to stay inside the request image caps the harness drops the oldest tool images. From that turn on the model no longer saw the message's images; the text it saw instead was the original text followed by a new line and [Older tool images omitted from active context; use view_image to inspect again.]. Null: never elided.
Each message's content is the version the model first saw. The tool results after the last assistant turn (normally the submit confirmation) were never sent back to the model.
harness_events (a JSON string): events the model did not see as messages (api_retry, unsafe_output_rejected, forced submits, ...), each with before_turn and detail.
tools and generation come from the first request; the export's checks confirm they never change within these rollouts.

Dropped: the raw request payloads (the whole growing context, resent every turn), opaque reasoning blobs (Anthropic thinking signatures, OpenAI encrypted reasoning content) and provider routing metadata.

Caveats

One rollout per scene and cohort. Rollouts of the same model and harness vary a lot (a repeated scene can move by tens of pixels), so single-scene differences are noisy.
iter01b was designed after analysing v0 rollouts. Its harness change was developed on t_dev_0001, t_test_0006, t_test_0047, t_test_0052, and t_test_0052 was examined in detail, so its gain on those scenes is partly in-sample (see the second table).
No ground-truth tie-point coordinates are included, only per-point errors (sn9_err_px) in tie-point order.
Host paths and account names in all text are replaced by $REGBENCH_ROOT, $WORK, $HOME, $USER and $GROUP; there the text differs from what the model saw. Account balances are removed.
The images are renders of SpaceNet-9 and BRIGHT imagery, and those datasets' licenses apply to them. Model outputs are subject to the providers' terms of use.

Downloads last month: 8

Total file size:

3 GB