Dataset Viewer
Auto-converted to Parquet Duplicate
run
stringclasses
8 values
sample
stringlengths
69
127
trace
stringlengths
80
127
task_id
stringclasses
84 values
upstream_task_id
stringclasses
28 values
run_index
int64
1
3
bench
stringclasses
1 value
category
stringclasses
5 values
model
stringclasses
8 values
harness
stringclasses
2 values
status
stringclasses
3 values
error
stringclasses
7 values
score
float64
0
1
⌀
grader_error
stringclasses
2 values
steps
int64
1
146
tool_calls
int64
0
148
input_tokens
int64
0
10.9M
output_tokens
int64
0
125k
total_tokens
int64
0
11M
elapsed_seconds
float64
26.7
1.28k
files
listlengths
0
100
messages
stringlengths
29.8k
667k
⌀
tools
stringclasses
2 values
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000000_01_Productivity_Flow_task_10_pdf_digest__run_01
"traces/codex-deepseek-v4.1-flash-1017/00000000_01_Productivity_Flow_task_10_pdf_digest__run_01.json(...TRUNCATED)
01_Productivity_Flow_task_10_pdf_digest__run_01
01_Productivity_Flow_task_10_pdf_digest
1
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0.9608
null
44
82
3,102,710
48,690
3,151,400
469.594
["paper_digest.md","papers/Asymmetric_Idiosyncrasies_in_Multimodal_Models.pdf","papers/BetterScene:_(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000001_01_Productivity_Flow_task_10_pdf_digest__run_02
"traces/codex-deepseek-v4.1-flash-1017/00000001_01_Productivity_Flow_task_10_pdf_digest__run_02.json(...TRUNCATED)
01_Productivity_Flow_task_10_pdf_digest__run_02
01_Productivity_Flow_task_10_pdf_digest
2
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0.9558
null
47
64
3,819,971
55,960
3,875,931
399.539
["paper_digest.md","papers/Asymmetric_Idiosyncrasies_in_Multimodal_Models.pdf","papers/BetterScene:_(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000002_01_Productivity_Flow_task_10_pdf_digest__run_03
"traces/codex-deepseek-v4.1-flash-1017/00000002_01_Productivity_Flow_task_10_pdf_digest__run_03.json(...TRUNCATED)
01_Productivity_Flow_task_10_pdf_digest__run_03
01_Productivity_Flow_task_10_pdf_digest
3
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0.9581
null
47
57
2,992,589
48,526
3,041,115
449.661
["paper_digest.md","papers/Asymmetric_Idiosyncrasies_in_Multimodal_Models.pdf","papers/BetterScene:_(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000003_01_Productivity_Flow_task_6_calendar_scheduling__run_01
"traces/codex-deepseek-v4.1-flash-1017/00000003_01_Productivity_Flow_task_6_calendar_scheduling__run(...TRUNCATED)
01_Productivity_Flow_task_6_calendar_scheduling__run_01
01_Productivity_Flow_task_6_calendar_scheduling
1
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
1
null
14
18
592,268
45,022
637,290
286.209
[ "decision_log.md", "scheduled.ics", "unscheduled.json" ]
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000004_01_Productivity_Flow_task_6_calendar_scheduling__run_02
"traces/codex-deepseek-v4.1-flash-1017/00000004_01_Productivity_Flow_task_6_calendar_scheduling__run(...TRUNCATED)
01_Productivity_Flow_task_6_calendar_scheduling__run_02
01_Productivity_Flow_task_6_calendar_scheduling
2
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0
null
18
24
818,635
40,107
858,742
244.832
[ "decision_log.md", "scheduled.ics", "unscheduled.json" ]
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000005_01_Productivity_Flow_task_6_calendar_scheduling__run_03
"traces/codex-deepseek-v4.1-flash-1017/00000005_01_Productivity_Flow_task_6_calendar_scheduling__run(...TRUNCATED)
01_Productivity_Flow_task_6_calendar_scheduling__run_03
01_Productivity_Flow_task_6_calendar_scheduling
3
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0
null
14
17
665,625
42,220
707,845
272.102
[ "decision_log.md", "scheduled.ics", "unscheduled.json" ]
null
null
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000006_01_Productivity_Flow_task_8_real_image_category__run_01
"traces/codex-deepseek-v4.1-flash-1017/00000006_01_Productivity_Flow_task_8_real_image_category__run(...TRUNCATED)
01_Productivity_Flow_task_8_real_image_category__run_01
01_Productivity_Flow_task_8_real_image_category
1
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0.976
null
55
107
3,704,327
53,420
3,757,747
496.526
["1_charts_and_tables/3d_rendered_interior.png","1_charts_and_tables/ancient_manuscript_page.jpg","1(...TRUNCATED)
null
null
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000007_01_Productivity_Flow_task_8_real_image_category__run_02
"traces/codex-deepseek-v4.1-flash-1017/00000007_01_Productivity_Flow_task_8_real_image_category__run(...TRUNCATED)
01_Productivity_Flow_task_8_real_image_category__run_02
01_Productivity_Flow_task_8_real_image_category
2
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0.968
null
22
41
946,169
37,185
983,354
294.885
["1_charts_and_tables/3d_rendered_interior.png","1_charts_and_tables/ancient_manuscript_page.jpg","1(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000008_01_Productivity_Flow_task_8_real_image_category__run_03
"traces/codex-deepseek-v4.1-flash-1017/00000008_01_Productivity_Flow_task_8_real_image_category__run(...TRUNCATED)
01_Productivity_Flow_task_8_real_image_category__run_03
01_Productivity_Flow_task_8_real_image_category
3
wildclawbench
01_Productivity_Flow
deepseek-v4.1-flash-1017
codex
ok
null
0.984
null
28
45
1,128,330
37,922
1,166,252
280.821
["1_charts_and_tables/3d_rendered_interior.png","1_charts_and_tables/ancient_manuscript_page.jpg","1(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
codex-deepseek-v4.1-flash-1017
samples/deepseek-v4.1-flash-1017/00000009_02_Code_Intelligence_task_10_acad_homepage_zh__run_01
"traces/codex-deepseek-v4.1-flash-1017/00000009_02_Code_Intelligence_task_10_acad_homepage_zh__run_0(...TRUNCATED)
02_Code_Intelligence_task_10_acad_homepage_zh__run_01
02_Code_Intelligence_task_10_acad_homepage_zh
1
wildclawbench
02_Code_Intelligence
deepseek-v4.1-flash-1017
codex
ok
null
0
null
58
108
3,342,047
65,650
3,407,697
560.5
["README.md","assets/IBMPlexMono-Regular.woff2","assets/SourceSansPro-Bold.woff2","assets/SourceSans(...TRUNCATED)
"[{\"role\": \"system\", \"content\": \"You are a coding agent running in the Codex CLI, a terminal-(...TRUNCATED)
"[{\"type\": \"function\", \"name\": \"exec_command\", \"description\": \"Runs a command in a PTY, r(...TRUNCATED)
End of preview. Expand in Data Studio

WildClawBench Trajectories

Graded agent trajectories of open-weight models, with every reasoning step and every tool call

Collected with Roller PyPI version of roller-agent 7 Harnesses 5 benchmarks

Roller on GitHub · How this dataset was made · Getting started · Documentation

1,008 trajectories across 28 WildClawBench tasks. Four model scores across OpenClaw, Codex and Claude Code.

1,008 trajectories. 28 tasks. Four models. Three Harnesses. 29,203 tool calls. DeepSeek-V4.1-Flash-1017, GLM-5.3-Flash, Kimi K3 and Kimi K3.1 Pro each run the same 28 WildClawBench tasks three times on OpenClaw, Codex and Claude Code. Every trajectory runs in its own Docker container, from the first prompt to the benchmark grader.

The dataset contains 336 trajectories per Harness. It includes successful, timed-out and errored runs.

The whole dataset is the output of Roller. One command collected and graded the trajectories, one exported them in this layout, and one uploaded them. The same three commands work with your model.

Trajectories Harness Model Samples Completed Scored Mean score Mean steps
openclaw-deepseek-v4.1-flash-1017 openclaw DeepSeek-V4.1-Flash 84 72 84 0.599 57.6
openclaw-glm-5.3-flash openclaw GLM-5.3-Flash 84 57 83 0.496 25.1
openclaw-roller-kimi-k3-dspark openclaw Kimi K3 84 61 81 0.489 15.8
openclaw-roller-kimi-k3.1-pro-dflash5 openclaw Kimi K3.1 Pro 84 75 80 0.575 16.6
codex-deepseek-v4.1-flash-1017 codex DeepSeek-V4.1-Flash 84 81 84 0.563 34.5
codex-glm-5.3-flash codex GLM-5.3-Flash 84 57 83 0.403 42.4
codex-kimi-k3 codex Kimi K3 84 39 74 0.334 15.8
codex-kimi-k3.1-pro codex Kimi K3.1 Pro 84 84 70 0.210 10.0
claude-code-deepseek-v4.1-flash-1017 claude-code DeepSeek-V4.1-Flash-1017 84 73 84 0.591 32.0
claude-code-glm-5.3-flash claude-code GLM-5.3-Flash 84 64 84 0.482 26.3
claude-code-kimi-k3 claude-code Kimi K3 84 39 75 0.354 13.0
claude-code-kimi-k3.1-pro claude-code Kimi K3.1 Pro 84 73 83 0.549 18.2

The mean score divides by all 84 samples of each model; missing scores count as zero. Model references retain their collection-time service names; the two roller-kimi-* names are Kimi K3 with DSpark and Kimi K3.1 Pro with DFLASH, respectively.

One trajectory, start to finish

DeepSeek-V4.1-Flash-1017 handles 15 meeting requests against a calendar of 30 existing events. The trajectory takes 33 model steps, 34 tool calls and 382 seconds; the grader scores it 1.0. This is an abridged extract from traces/openclaw-deepseek-v4.1-flash-1017/00000003_01_Productivity_Flow_task_6_calendar_scheduling__run_01.jsonl:

user       I manage a busy team calendar and need help resolving scheduling conflicts.
tool       exec {"command":"ls -la /tmp_workspace/ && echo \"---\" && cat /tmp_workspace/constraints.yaml"}
tool       exec {"command":"cat /tmp_workspace/meeting_requests.json"}
tool       exec {"command":"cat /tmp_workspace/calendar.ics"}
assistant  Completed the schedule optimization and saved all three required files in `/tmp_workspace/results/`.
grader     score 1.0 · decision_log.md, scheduled.ics, unscheduled.json

The native sessions preserve the recorded prompt, reasoning, tool calls and tool results. The table adds status, output file names, score and grader error. Open Data Studio and click a row to read one in the trace viewer of the Hub.

Three commands

The OpenClaw recipe was run separately for each model. Set MODEL to a model reference from the collection settings below, and provide the matching entry in models.json. All runs used the same task list, image, output budget and benchmark timeouts.

git clone https://github.com/Valen-Team/Roller.git
cd Roller
pip install -e .
cd examples/cookbook/wildclawbench-open-models

export WILDCLAWBENCH_ROOT=~/WildClawBench
export OPENROUTER_BASE_URL=http://localhost:8001/v1
export OPENROUTER_API_KEY=not-required
export JUDGE_MODEL=glm-5.3-flash
MODEL=vllm/glm-5.3-flash

# 1. Collect and grade. One Docker container per trajectory, graded before the container is removed.
roller collect --bench wildclawbench --bench-root "$WILDCLAWBENCH_ROOT" \
  $(sed 's/^/--bench-task-id /' task-ids.txt) --bench-runs-per-task 3 \
  --model "$MODEL" \
  --model-config models.json \
  --harness openclaw --openclaw-image ghcr.io/valen-team/roller-wildclaw-base:1.3-clean.1 \
  --openclaw-max-tokens 32768 --task-timeouts --keep-thinking --wire-trace full \
  --parallel 8 --grader-parallel 2 \
  --run-name wildclawbench-open-models

# 2. Export as a Hugging Face dataset: traces/, data/ and the results table of this card.
roller export runs/wildclawbench-open-models --format hf --output dataset \
  --preserve-task-keys --redact-rules redact.json

# 3. Upload the folder as it is.
hf upload Valen-Team/Roller-WildClawBench-Trajectories dataset . --repo-type dataset

The Codex batch uses the same task list and judge. Its runnable script selects the Codex image and Responses settings:

export MODEL=glm-5.3-flash
export OPENAI_BASE_URL=http://localhost:8000/v1
export OPENAI_API_KEY=not-required
export MODEL_CONTEXT_WINDOW=98304
export RUN_NAME=wildclawbench-codex-glm
./collect-codex.sh --dry-run
./collect-codex.sh
roller export runs/wildclawbench-codex-glm --format hf --output dataset \
  --preserve-task-keys --redact-rules redact.json

Use your own endpoints, authentication and serving context window. These new options require the Git installation above; they are not in PyPI 0.4.0. The task images and their build instructions remove personal builder credentials and identity state while keeping task assets and deliberate benchmark keys unchanged.

The export replaces what is private to the collection environment before anything leaves the machine: model endpoints become endpoint-<n>.example, the hosts of NO_PROXY become internal-<n>.example, the name of the machine becomes collection-host, host paths become placeholders such as <bench-root>, and credential values become [REDACTED]. redact.json adds what the run does not record, here the directory of the model weights.

Change one argument to collect your own data:

To change Argument
The model --model, with an entry in models.json for any OpenAI-compatible endpoint: vLLM, SGLang, or a hosted API
The agent Harness --harness codex, claude-code, hermes, opencode, deepseek, or pi
The benchmark --bench skillsbench, pinchbench, qwenclawbench, or swebench-pro
The tasks --tasks tasks.jsonl, with one prompt and its input files per line

The cookbook recipe walks through this run step by step, and examples/cookbook/wildclawbench-open-models holds its models.json, task list, and script.

The published runs were collected separately: 8 rollout workers per model, with the concurrent GLM/Kimi jobs sharing two grader slots across the jobs. The single-job command above uses two graders for that job.

Why collect with Roller

  • A benchmark is one argument. --bench converts the benchmark tasks, runs them, and grades each trajectory in the container where it ran. Five benchmarks are built in, and your own tasks are a JSONL file.
  • Seven agent Harnesses, one task format. OpenClaw, Codex, Claude Code, Hermes, OpenCode, DeepSeek Harness, and Pi Agent write the same sample layout, so trajectories from different agents land in one dataset.
  • Every run stays visible. The table covers all 1,008 attempts, with 1,008 native sessions. Chat-format messages and tools are available for 749 samples.
  • Built for batches. Rollouts run in parallel, failed ones retry, and --resume continues an interrupted run.
  • Publishing is one command. roller export --format hf writes the files that the trace viewer of the Hub reads, the table that datasets loads, and the results table of the card.
  • A Viewer for your own runs. roller view opens a run directory in the browser: search the tasks, compare models on one task, and read each step.

Roller Viewer stepping through the reasoning, tool calls and tool results of one rollout

Use the dataset

Load the table. It has one row per trajectory, and messages and tools are JSON strings:

import json

from datasets import load_dataset

rows = load_dataset(
    "Valen-Team/Roller-WildClawBench-Trajectories", "trajectories", split="train"
)
row = rows[0]
print(row["run"], row["upstream_task_id"], row["score"], row["steps"])
messages = json.loads(row["messages"])

Keep the trajectories that scored at least 0.8 as training records. Assistant messages carry reasoning_content:

solved = rows.filter(
    lambda row: row["messages"] is not None and (row["score"] or 0) >= 0.8
)
records = [
    {"messages": json.loads(row["messages"]), "tools": json.loads(row["tools"])}
    for row in solved
]

Compare the models task by task:

frame = rows.to_pandas()
print(frame.pivot_table(index="upstream_task_id", columns="run", values="score"))

Layout

Every path carries a <harness>-<model> name, such as openclaw-glm-5.3-flash, so each model and Harness has its own part.

Path Content
traces/<harness>-<model>/<index>_<task>.jsonl The session of one trajectory, as the agent Harness wrote it. The traces configuration, which the Hub opens in its trace viewer
data/<harness>-<model>.jsonl One row per trajectory. The trajectories configuration

In an OpenClaw session, Roller writes four fields to the session header and leaves every other line as OpenClaw wrote it: harness, a session ID that is unique in this dataset, the session name, and a roller object with the <harness>-<model> name, sample path, task, status, and score.

Table columns

Column Content
run, sample, trace The <harness>-<model> name, the sample path of the trajectory in the Roller run, and the path of its session under traces/
task_id, upstream_task_id, run_index Task ID of this repetition, the WildClawBench task ID, and the repetition number
bench, category wildclawbench and the WildClawBench category
model, harness Model reference and agent Harness
status, error ok for a trajectory that finished, otherwise error or timeout with the reason
score, grader_error WildClawBench score from 0 to 1. score is empty when grading failed, and grader_error gives the reason
steps, tool_calls Number of model steps and of tool calls
input_tokens, output_tokens, total_tokens, elapsed_seconds Usage and duration of the trajectory
files Names of the files the agent wrote to the task output directory
messages, tools The trajectory as an OpenAI chat request, as two JSON strings. Assistant messages carry reasoning_content. Available where a chat-format conversion was recorded

OpenClaw collections

The four collections contain 84 trajectories each: 28 tasks × 3 repetitions. The two Kimi service names retain the deployment suffixes used during collection.

Item Value
Benchmark WildClawBench, revision 41bfe78a4d04369887687397708504882a80c17a
Tasks 28 of 60, across Productivity Flow, Code Intelligence, Social Interaction, Creative Synthesis and Safety Alignment
Repetitions 3 per task and model
Harness OpenClaw 2026.3.11, image wildclawbench-ubuntu:v1.3
Roller collection 034fdaffbbe9f327618a51e5f04a9b49dec875a1
Parallelism 8 rollouts per model; 2 graders. Concurrent GLM/Kimi runs shared a global 2-slot grader budget
Timeouts Original WildClawBench task timeouts, multiplier 1; grader timeout 120 seconds
Retries No retries of normally ended failed or timed-out samples
Output budget 32,768 tokens per response, thinking retained, full wire traces captured in the private run
Network web_search disabled; external network access limited by the job environment. Warmup dependencies staged as offline wheels
Judge GLM-5.3-Flash for all four models, invoked inside each task container
Collected 16 November 2026, Asia/Singapore
Model Collection model reference Serving and reasoning
DeepSeek-V4.1-Flash-1017 vllm/deepseek-v4.1-flash-1017 vLLM, TP4, DSpark=5, Engram CPU offload; thinking=true, reasoning_effort=100
GLM-5.3-Flash vllm/glm-5.3-flash vLLM, TP4, MTP=5; reasoning_effort=max
Kimi K3 vllm/roller-kimi-k3-dspark vLLM, TP16/EP16, DSpark=8; thinking_effort=max
Kimi K3.1 Pro vllm/roller-kimi-k3.1-pro-dflash5 SGLang, TP16/EP16, DFLASH width 5, Marlin MoE; default reasoning_effort=max

Reasoning settings come from the recorded server configuration. The collection client did not explicitly set temperature or top_p; sampling follows the serving engine's defaults. DeepSeek ran first (19:50:13–21:18:02), then GLM (21:37:07–23:24:03), Kimi K3 (21:55:07–23:39:59) and Kimi K3.1 Pro (22:02:30–23:26:11), all on 16 November in Asia/Singapore. The end times include grading and run finalization.

Graders check local files and tool calls, or call the GLM-5.3-Flash judge.

Model ok Timeout Error Missing valid score
DeepSeek-V4.1-Flash 72 10 2 0
GLM-5.3-Flash 57 22 5 1
Kimi K3 61 21 2 3
Kimi K3.1 Pro 75 9 0 4

All 336 records are retained. Of these, 328 have valid scores. Eight lack required grader inputs (screenshot.png or results.md); their score is null and grader_error explains why. A timeout or runtime error can still have a score if the benchmark grader can inspect the partial result.

DeepSeek and GLM Codex collections

Added on 17 November 2026: DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash each run the same 28 tasks three times, adding 168 trajectories.

Item Value
Harness Codex CLI 0.149.0
Task image ghcr.io/valen-team/roller-wildclaw-codex:1.3-codex0.149.0
Roller collection revision a392cc0
Judge GLM-5.3-Flash
Parallelism 8 rollouts per model, shared 2-slot grader budget
Timeouts Original task limits; 120-second grader timeout
Retries 0 agent retries and 0 grader retries
Output budget 32,768 tokens per response; recorded backend thinking defaults retained
Serving DeepSeek: DSpark=5, thinking enabled, effort 100; GLM: MTP=5, effort max
Model ok Timeout Error Missing valid score
DeepSeek-V4.1-Flash 81 3 0 0
GLM-5.3-Flash 57 27 0 1

All 168 attempts are retained; 167 have valid scores. The missing score remains null with its grader error. Raw run directories and wire captures stay in the collection environment.

Kimi Codex collections

Kimi K3 and Kimi K3.1 Pro each run the same 28 WildClawBench tasks three times with Codex CLI 0.149.0 and the GLM-5.3-Flash judge. The original task prompts and timeout limits are used for all 168 attempts. The coverage counts below include all three attempts per task.

Kimi K3 uses vLLM 0.33.0, which applies the Kimi parser's special-token spacing setting to streaming Responses tool arguments. Kimi K3.1 Pro uses the same SGLang tool-image compatibility mode as the initial collection.

Model ok Timeout Error Missing valid score
Kimi K3 39 45 0 10
Kimi K3.1 Pro 84 0 0 14

A missing score appears as null in the trajectory table, with the grader's error when available.

Claude Code collections

Claude Code 2.1.222 ran the same 28 tasks three times per model with the GLM-5.3-Flash judge and the original task timeouts. The collection script and task image reproduce the setup with an Anthropic Messages endpoint.

Model Completed Timeout Error Missing score
DeepSeek-V4.1-Flash-1017 73 11 0 0
GLM-5.3-Flash 64 15 5 0
Kimi K3 39 39 6 9
Kimi K3.1 Pro 73 11 0 1

The four endpoints used Anthropic Messages. Claude Code requested up to 32,000 output tokens. The collection used pinned Claude Code 2.1.222 and eight rollout workers per model, with two judge slots shared across the jobs.

The affected PDF requests were rerun using Roller's --claude-pdf-as-images option, which forwards extracted text and a PNG for each PDF page. For the affected GLM-5.3-Flash runs, Claude Code was configured with the endpoint's 98,304-token context window. Five reruns still exceeded that limit and their error status is retained. Each table contains one row per scheduled task and repetition.

License

The dataset is MIT licensed, with attribution to WildClawBench. The model cards linked above provide the licenses for each model.

Citation

@misc{roller,
  author       = {{Valen Team}},
  title        = {Roller},
  howpublished = {GitHub - Valen-Team/Roller: Agent trajectory collection across models and Harnesses.},
  note         = {GitHub repository},
  year         = {2026},
  url          = {https://github.com/Valen-Team/Roller}
}

@article{ding2026wildclawbench,
  title={WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation},
  author={Ding, Shuangrui and Dai, Xuanlang and Xing, Long and Ding, Shengyuan and Liu, Ziyu and JingYi, Yang and Yang, Penghui and Zhang, Zhixiong and Wei, Xilin and Fang, Xinyu and others},
  journal={arXiv preprint arXiv:2605.10912},
  year={2026}
}
Downloads last month
238

Paper for Valen-Team/Roller-WildClawBench-Trajectories