Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
run: string
direction: string
instance_id: string
arm: string
arm_id: string
model_alias: string
fork_step: int64
exit_status: string
n_steps: int64
prompt_tokens: int64
completion_tokens: int64
replay_mismatches: int64
wall_time_s: double
resolved: bool
patch_len: int64
messages: null
patch: string
replay_fidelity: list<item: struct<action_index: int64, command: string, replayed_returncode: int64, recorded_returnc (... 25 chars omitted)
child 0, item: struct<action_index: int64, command: string, replayed_returncode: int64, recorded_returncode: int64, (... 13 chars omitted)
child 0, action_index: int64
child 1, command: string
child 2, replayed_returncode: int64
child 3, recorded_returncode: int64
child 4, match: bool
to
{'run': Value('string'), 'direction': Value('string'), 'instance_id': Value('string'), 'arm': Value('string'), 'arm_id': Value('string'), 'model_alias': Value('string'), 'fork_step': Value('int64'), 'exit_status': Value('string'), 'patch': Value('string'), 'n_steps': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'replay_mismatches': Value('int64'), 'replay_fidelity': List({'action_index': Value('int64'), 'command': Value('string'), 'replayed_returncode': Value('int64'), 'recorded_returncode': Value('int64'), 'match': Value('bool')}), 'wall_time_s': Value('float64'), 'resolved': Value('bool'), 'messages': List(Json(decode=True))}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1816, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
run: string
direction: string
instance_id: string
arm: string
arm_id: string
model_alias: string
fork_step: int64
exit_status: string
n_steps: int64
prompt_tokens: int64
completion_tokens: int64
replay_mismatches: int64
wall_time_s: double
resolved: bool
patch_len: int64
messages: null
patch: string
replay_fidelity: list<item: struct<action_index: int64, command: string, replayed_returncode: int64, recorded_returnc (... 25 chars omitted)
child 0, item: struct<action_index: int64, command: string, replayed_returncode: int64, recorded_returncode: int64, (... 13 chars omitted)
child 0, action_index: int64
child 1, command: string
child 2, replayed_returncode: int64
child 3, recorded_returncode: int64
child 4, match: bool
to
{'run': Value('string'), 'direction': Value('string'), 'instance_id': Value('string'), 'arm': Value('string'), 'arm_id': Value('string'), 'model_alias': Value('string'), 'fork_step': Value('int64'), 'exit_status': Value('string'), 'patch': Value('string'), 'n_steps': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'replay_mismatches': Value('int64'), 'replay_fidelity': List({'action_index': Value('int64'), 'command': Value('string'), 'replayed_returncode': Value('int64'), 'recorded_returncode': Value('int64'), 'match': Value('bool')}), 'wall_time_s': Value('float64'), 'resolved': Value('bool'), 'messages': List(Json(decode=True))}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
run string | direction string | instance_id string | arm string | arm_id string | model_alias string | fork_step int64 | exit_status string | patch string | n_steps int64 | prompt_tokens int64 | completion_tokens int64 | replay_mismatches int64 | replay_fidelity list | wall_time_s float64 | resolved bool | messages list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
easy | up | astropy__astropy-7166 | base | base | null | null | LimitsExceeded | 35 | 316,740 | 5,428 | null | null | 165.372388 | false | [
{
"role": "system",
"content": "You are a helpful assistant that can interact with a computer shell to solve programming tasks."
},
{
"role": "user",
"content": "<pr_description>\nConsider the following PR description:\nInheritDocstrings metaclass doesn't work for properties\nInside the InheritD... | |
easy | up | astropy__astropy-7166 | branch | large_k10 | large | 10 | Submitted | "diff --git a/astropy/utils/misc.py b/astropy/utils/misc.py\nindex de3fb01d1..9aab94e15 100644\n--- (...TRUNCATED) | 12 | 44,541 | 3,410 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"inherit\"","replayed_returnc(...TRUNCATED) | 60.665419 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) |
easy | up | astropy__astropy-7166 | branch | large_k24 | large | 24 | Submitted | "diff --git a/astropy/utils/misc.py b/astropy/utils/misc.py\nindex de3fb01d1..1b1cb22c9 100644\n--- (...TRUNCATED) | 26 | 169,323 | 3,615 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"inherit\"","replayed_returnc(...TRUNCATED) | 8.510715 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) |
easy | up | astropy__astropy-7166 | branch | small_k10 | small | 10 | LimitsExceeded | 33 | 349,088 | 5,721 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"inherit\"","replayed_returnc(...TRUNCATED) | 139.554408 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) | |
easy | up | astropy__astropy-7166 | branch | small_k24 | small | 24 | Submitted | "diff --git a/astropy/utils/misc.py b/astropy/utils/misc.py\nindex de3fb01d1..1b1cb22c9 100644\n--- (...TRUNCATED) | 32 | 269,516 | 5,010 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"inherit\"","replayed_returnc(...TRUNCATED) | 51.424046 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) |
easy | up | astropy__astropy-7336 | base | base | null | null | LimitsExceeded | 27 | 291,471 | 4,247 | null | null | 129.602828 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) | |
easy | up | astropy__astropy-7336 | branch | large_k19 | large | 19 | LimitsExceeded | 47 | 415,458 | 4,891 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"decorators\"","replayed_retu(...TRUNCATED) | 58.114155 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) | |
easy | up | astropy__astropy-7336 | branch | large_k8 | large | 8 | Submitted | "diff --git a/astropy/units/decorators.py b/astropy/units/decorators.py\nindex 8bece5a85..9044b35a7 (...TRUNCATED) | 9 | 39,452 | 1,337 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"decorators\"","replayed_retu(...TRUNCATED) | 3.52186 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) |
easy | up | astropy__astropy-7336 | branch | small_k19 | small | 19 | LimitsExceeded | 34 | 423,624 | 5,339 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"decorators\"","replayed_retu(...TRUNCATED) | 74.781331 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) | |
easy | up | astropy__astropy-7336 | branch | small_k8 | small | 8 | LimitsExceeded | 29 | 320,295 | 4,648 | 0 | [{"action_index":0,"command":"find . -type f -name \"*.py\" | grep -i \"decorators\"","replayed_retu(...TRUNCATED) | 103.306547 | false | [{"role":"system","content":"You are a helpful assistant that can interact with a computer shell to (...TRUNCATED) |
End of preview.
The Replay Gap: Branched Agent Trajectories
Counterfactual ("branched") agent rollouts for studying per-step model switching in LLM agents, from the paper The Replay Gap: Static Evaluation of Model Switching in LLM Agents Scores the Wrong World (Efficient Reasoning Workshop @ COLM 2026).
Routing benchmarks score routers by replaying logged model outputs. In a multi-step agent that is unsound: swap the model at step k and the rest of the trajectory diverges. This dataset contains the ground truth replay cannot provide — trajectories actually forked mid-run and continued live with a different model, each paired with a same-model control fork that isolates sampling and environment-replay noise.
What's in it
~900 containerized SWE-bench Verified rollouts across six seed-matched run pairs (difficulty tier × swap direction):
| run | base model | branch arms | tier |
|---|---|---|---|
pilot30 / pilot30_rev |
4B / 14B | both models | full difficulty |
easy / easy_rev |
4B / 14B | both models | "<15 min fix" |
nudge / nudge_rev |
4B / 14B | both models | "<15 min fix", budget-nudged prompt |
Each instance contributes one base trajectory plus branches forked at 30% and 70% of its length, each continued by both the same model (control) and the other model (swap).
Fields
| field | description |
|---|---|
run, direction |
run pair and swap direction (up = 4B base, down = 14B base) |
instance_id |
SWE-bench Verified instance |
arm, arm_id, model_alias, fork_step |
base vs branch; which model continued; the assistant turn forked at |
messages |
full message history (system, user, assistant turns with parsed actions, observations) |
exit_status |
Submitted, LimitsExceeded, ContextWindowExceededError, RepeatedFormatError |
patch |
submitted git patch (may be empty) |
resolved |
official SWE-bench harness outcome, null if that arm was not scored |
replay_fidelity, replay_mismatches |
per-action return-code agreement when rebuilding the pre-fork environment |
n_steps, prompt_tokens, completion_tokens, wall_time_s |
cost and length |
rollouts_index.jsonl is the same data without message bodies, for quick
filtering.
Generation setup
mini-SWE-agent (bash-only ReAct loop) on official SWE-bench Verified Docker images; 50-step budget, 28k context. Models served by vLLM at temperature 0 on a single 24GB GPU: Qwen3-4B-Instruct-2507-FP8 and Qwen3-14B-AWQ (thinking disabled). Prefix replay fidelity: 99.99% return-code agreement across 11,702 replayed actions; 707/708 branches reconstructed exactly.
Caveats
Absolute task-resolution rates are low (0–3%) because of the constrained serving budget — the action-level signal is robust, but outcome-level analyses rest on five flip events. The two models differ in both scale and quantization stack (FP8 vs AWQ), so up/down comparisons are between deployment configurations, not pure model scales. One scaffold, one benchmark family.
Citation
@inproceedings{gonuguntla2026replaygap,
title = {The Replay Gap: Static Evaluation of Model Switching in {LLM} Agents Scores the Wrong World},
author = {Gonuguntla, Ashritha},
booktitle = {Efficient Reasoning Workshop at COLM},
year = {2026},
url = {https://openreview.net/forum?id=8gqqiNrzyA}
}
Code: https://github.com/AshrithaG/replay-gap Project page: https://ashrithag.github.io/replay-gap/
- Downloads last month
- 83