Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
status: string
passed: bool
policy: string
start_mode: string
source_input: string
source_input_sha256: string
source_roots: int64
eligible_roots: int64
retained_requests: int64
workload_sha256: string
maximum_request: struct<root: string, request: string, input: int64, output: int64, total: int64>
child 0, root: string
child 1, request: string
child 2, input: int64
child 3, output: int64
child 4, total: int64
tokenizer: struct<path: string, tokenizer_json_sha256: string>
child 0, path: string
child 1, tokenizer_json_sha256: string
model_config_sha256: string
roots: list<item: struct<id: string, requests: int64, source_row_sha256: string>>
child 0, item: struct<id: string, requests: int64, source_row_sha256: string>
child 0, id: string
child 1, requests: int64
child 2, source_row_sha256: string
checks: list<item: string>
child 0, item: string
limitation: string
verdict: string
attests: struct<workload_sha256: string, tokenizer_sha256: string>
child 0, workload_sha256: string
child 1, tokenizer_sha256: string
generates: int64
sources: list<item: string>
child 0, item: string
prompt_tokens: int64
error_count: int64
traces: int64
sessions: list<item: struct<chain_overlaps: int64, chained_pairs_checked: int64, chains: struct<derived: int64 (... 373 chars omitted)
child 0, item: struct<chain_overlaps: int64, chained_pairs_checked: int64, chains: struct<derived: int64, main: int (... 361 chars omitted)
child 0, chain_overlaps: int64
child 1, chained_pairs_checked: int64
child 2, chains: struct<derived: int64, main: int64, subagent: int64>
child 0, derived: int64
child 1, main: int64
child 2, subagent: int64
child 3, join_modes: struct<derived:background: int64, derived:end: int64, subagent:blocking: int64, derived:blocking: in (... 53 chars omitted)
child 0, derived:background: int64
child 1, derived:end: int64
child 2, subagent:blocking: int64
child 3, derived:blocking: int64
child 4, subagent:background: int64
child 5, subagent:end: int64
child 4, output_tokens: int64
child 5, prompt_tokens: int64
child 6, requests: int64
child 7, seams_merged: int64
child 8, session: string
child 9, session_prefix_hit_blocks: struct<hash: int64, of: int64, prompt: int64>
child 0, hash: int64
child 1, of: int64
child 2, prompt: int64
workload: string
errors: list<item: null>
child 0, item: null
output_tokens: int64
to
{'attests': {'tokenizer_sha256': Value('string'), 'workload_sha256': Value('string')}, 'error_count': Value('int64'), 'errors': List(Value('null')), 'generates': Value('int64'), 'output_tokens': Value('int64'), 'prompt_tokens': Value('int64'), 'sessions': List({'chain_overlaps': Value('int64'), 'chained_pairs_checked': Value('int64'), 'chains': {'derived': Value('int64'), 'main': Value('int64'), 'subagent': Value('int64')}, 'join_modes': {'derived:background': Value('int64'), 'derived:end': Value('int64'), 'subagent:blocking': Value('int64'), 'derived:blocking': Value('int64'), 'subagent:background': Value('int64'), 'subagent:end': Value('int64')}, 'output_tokens': Value('int64'), 'prompt_tokens': Value('int64'), 'requests': Value('int64'), 'seams_merged': Value('int64'), 'session': Value('string'), 'session_prefix_hit_blocks': {'hash': Value('int64'), 'of': Value('int64'), 'prompt': Value('int64')}}), 'sources': List(Value('string')), 'traces': Value('int64'), 'verdict': Value('string'), 'workload': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
status: string
passed: bool
policy: string
start_mode: string
source_input: string
source_input_sha256: string
source_roots: int64
eligible_roots: int64
retained_requests: int64
workload_sha256: string
maximum_request: struct<root: string, request: string, input: int64, output: int64, total: int64>
child 0, root: string
child 1, request: string
child 2, input: int64
child 3, output: int64
child 4, total: int64
tokenizer: struct<path: string, tokenizer_json_sha256: string>
child 0, path: string
child 1, tokenizer_json_sha256: string
model_config_sha256: string
roots: list<item: struct<id: string, requests: int64, source_row_sha256: string>>
child 0, item: struct<id: string, requests: int64, source_row_sha256: string>
child 0, id: string
child 1, requests: int64
child 2, source_row_sha256: string
checks: list<item: string>
child 0, item: string
limitation: string
verdict: string
attests: struct<workload_sha256: string, tokenizer_sha256: string>
child 0, workload_sha256: string
child 1, tokenizer_sha256: string
generates: int64
sources: list<item: string>
child 0, item: string
prompt_tokens: int64
error_count: int64
traces: int64
sessions: list<item: struct<chain_overlaps: int64, chained_pairs_checked: int64, chains: struct<derived: int64 (... 373 chars omitted)
child 0, item: struct<chain_overlaps: int64, chained_pairs_checked: int64, chains: struct<derived: int64, main: int (... 361 chars omitted)
child 0, chain_overlaps: int64
child 1, chained_pairs_checked: int64
child 2, chains: struct<derived: int64, main: int64, subagent: int64>
child 0, derived: int64
child 1, main: int64
child 2, subagent: int64
child 3, join_modes: struct<derived:background: int64, derived:end: int64, subagent:blocking: int64, derived:blocking: in (... 53 chars omitted)
child 0, derived:background: int64
child 1, derived:end: int64
child 2, subagent:blocking: int64
child 3, derived:blocking: int64
child 4, subagent:background: int64
child 5, subagent:end: int64
child 4, output_tokens: int64
child 5, prompt_tokens: int64
child 6, requests: int64
child 7, seams_merged: int64
child 8, session: string
child 9, session_prefix_hit_blocks: struct<hash: int64, of: int64, prompt: int64>
child 0, hash: int64
child 1, of: int64
child 2, prompt: int64
workload: string
errors: list<item: null>
child 0, item: null
output_tokens: int64
to
{'attests': {'tokenizer_sha256': Value('string'), 'workload_sha256': Value('string')}, 'error_count': Value('int64'), 'errors': List(Value('null')), 'generates': Value('int64'), 'output_tokens': Value('int64'), 'prompt_tokens': Value('int64'), 'sessions': List({'chain_overlaps': Value('int64'), 'chained_pairs_checked': Value('int64'), 'chains': {'derived': Value('int64'), 'main': Value('int64'), 'subagent': Value('int64')}, 'join_modes': {'derived:background': Value('int64'), 'derived:end': Value('int64'), 'subagent:blocking': Value('int64'), 'derived:blocking': Value('int64'), 'subagent:background': Value('int64'), 'subagent:end': Value('int64')}, 'output_tokens': Value('int64'), 'prompt_tokens': Value('int64'), 'requests': Value('int64'), 'seams_merged': Value('int64'), 'session': Value('string'), 'session_prefix_hit_blocks': {'hash': Value('int64'), 'of': Value('int64'), 'prompt': Value('int64')}}), 'sources': List(Value('string')), 'traces': Value('int64'), 'verdict': Value('string'), 'workload': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Evolving Infra agent traces
The agent workloads of the Evolving Infra
evaluation. Each file is a workload for the replay client in
benchmarks/trie: every
request has a fixed prompt and a fixed output length, and requests keep the recorded agent
structure (sub-agents, tool waits, shared prefixes), so two serving systems see the same work.
| Workload | Agent | Tasks | Requests | Output tokens | File |
|---|---|---|---|---|---|
| Claude Code | Claude Code, real users | 393 | 98,786 | 106 M | claude-code/cc062126.portable.jsonl |
| Putnam | mini-SWE-agent on PutnamBench | 100 | 6,348 | 14.5 M | putnam/decode-heavy-100.stationary.jsonl |
| OSWorld | computer-use agent on OSWorld | 98 | 14,920 | 2.0 M | osworld/osworld2-kimi-formal.jsonl |
The replay client turns a workload into the served model's input:
| Workload | Stored as | On the served model |
|---|---|---|
| Claude Code | 64-token hash blocks | The client draws each block's tokens from the model's vocabulary (tokenizer_model=<model directory>, which needs tokenizer.json). |
| Putnam | messages (text) | The client renders the prompts with the model's chat template. |
| OSWorld | messages, screenshots, and Kimi-K2.6's prompt token IDs and image layout | Replays on Kimi-K2.6 as is. For another model, bind the messages and screenshots to its image layout first (benchmarks/trie/scripts/bind_cua_kimi_images.py, convert_cua_deepseek_v4.py). |
Each file has a .sha256 file with its checksum.
Claude Code
Built from SemiAnalysis's AgentX Claude Code traces (Apache-2.0), collected from real users. The source has prefix hashes, lengths and timing but no text, so prompts are stored as 64-token hash blocks. Prefix sharing between requests and sub-agents is kept. Tool and user waits longer than 120 s are clipped to 120 s. The prompts add up to 21.6 B tokens; the longest request has 987 K prompt tokens.
Putnam
mini-SWE-agent solving PutnamBench theorems in Lean. Each task keeps the recorded messages, tool calls and outputs; the agent was served by GLM-5.2 when it was recorded.
OSWorld
A computer-use agent working on OSWorld tasks through screenshots and normalized click and
keyboard tools, recorded with Kimi-K2.6. osworld/screenshots.tar holds the 12,838 screenshots
the requests reference (5.9 GB), named by sha256 (tar -xf osworld/screenshots.tar -C osworld).
The workload's metadata.image_root is osworld/screenshots, relative to the download
directory; --image-root in make_shards.py points it elsewhere.
Files of the paper's runs
The paper's Claude Code runs replayed the same workload with one model's vocabulary written into the file. With GLM-5.2's tokenizer, the portable file sends the token IDs of the GLM-5.2 file.
| File | Model | Runs | Offline check |
|---|---|---|---|
claude-code/cc062126-stationary-full-v1.jsonl |
GLM-5.2 | GB300 | .validation.json |
claude-code/cc062126-dsv4-flash.jsonl |
DeepSeek-V4-Flash (deepseek-ai/DeepSeek-V4-Flash-0731) |
H200 | .validation.json |
The Putnam and OSWorld runs used the files in the first table.
Use
hf download crmsndu/evolving-infra-traces --repo-type dataset --local-dir traces
cd traces
tar -xf osworld/screenshots.tar -C osworld # OSWorld only
The paper replays Claude Code with one client process and Putnam and OSWorld with 16.
shards.json lists which tasks each process reads, and make_shards.py writes those files:
python make_shards.py putnam/decode-heavy-100.stationary.jsonl putnam/shards.json shards/putnam
python make_shards.py osworld/osworld2-kimi-formal.jsonl osworld/shards.json shards/osworld \
--image-root "$PWD/osworld/screenshots"
Pass a workload or a shard to the replay client as workload_path, with tokenizer_model=<model directory>; see benchmarks/replay.
- Downloads last month
- 222