Datasets:

---

FWeindel

/

validated-tasks

Like 0

Traces and pass counts of the pass@4 run of mimo_v1 with qwen3-30b-instruct-2507 (Qwen/Qwen3-30B-A3B-Instruct-2507): 10036 attempts of 2509 tasks; 77 attempts solved, 59 tasks solved at least once; pass@1 0.8%, pass@4 2.4%.

• runs/2026-11-06-mimo_v1-pass-counts-adeb1412ff63/trajectories/mimo_v1/qwen3-30b-instruct-2507.parquet: one row per attempt with the trajectory, the verifier record (result.json) and the reward, as validation/publishing/trajectories.py writes them.
• mimo_v1/pass_counts.parquet: one row per task (path, content_sha256, model, attempts, graded, solved), like the other pass_counts.parquet files. Tasks by solved attempts: 0/4: 2450, 1/4: 46, 2/4: 9, 3/4: 3, 4/4: 1.

How it was run: on a local server (NVIDIA H100 NVL) with hpc/passk_local.sh from OT-next-data branch passk-script (commit 35c32330e9c1, stage layout environment-first-v2), not on a cluster. Docker instead of Apptainer, with the task images built on the server; 24 attempts at once on one GPU; the model served by vLLM with the settings of config/models.py as the cluster worker starts it, except --max-num-seqs 128; agent terminus-2; Harbor 7faf878c14b6. Network as each task.toml sets it: 2509 of 2509 tasks allow internet access.

Versions: tasks FWeindel/validated-tasks/mimo_v1 at a8acfde13cc1be1672da3d17ee70bc4d842b4c19; weights Qwen/Qwen3-30B-A3B-Instruct-2507 at 0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe; contract sha256 e4c6e2318a7f010af7ef0a1673a911f17d225c2aeeed0cebc1d9e68cd3009f57.

In the records, this server's paths, user name and host name are replaced by , , and ; nothing else was changed.

Traces and pass counts of the pass@4 run of mimo_v1 with qwen3-30b-thinking (Qwen/Qwen3-30B-A3B): 10036 attempts of 2509 tasks; 54 attempts solved, 43 tasks solved at least once; pass@1 0.5%, pass@4 1.7%.

• runs/2026-11-09-mimo_v1-pass-counts-1bfea1af4bc4/trajectories/mimo_v1/qwen3-30b-thinking.parquet: one row per attempt with the trajectory, the verifier record (result.json) and the reward, as validation/publishing/trajectories.py writes them.
• mimo_v1/pass_counts.parquet: one row per task and model (path, content_sha256, model, attempts, graded, solved), like the other pass_counts.parquet files, for Qwen/Qwen3-30B-A3B and Qwen/Qwen3-30B-A3B-Instruct-2507. Qwen/Qwen3-30B-A3B, tasks by solved attempts: 0/4: 2466, 1/4: 34, 2/4: 8, 3/4: 0, 4/4: 1.

How it was run: on a local server (NVIDIA H100 NVL) with hpc/passk_local.sh from OT-next-data branch passk-script (commit b0f474334125, stage layout environment-first-v2), not on a cluster. Docker instead of Apptainer, with the task images built on the server; 24 attempts at once on one GPU; the model served by vLLM with the settings of config/models.py as the cluster worker starts it, except --max-num-seqs 128; agent terminus-2; Harbor 7faf878c14b6. Network as each task.toml sets it: 2509 of 2509 tasks allow internet access.

Versions: tasks FWeindel/validated-tasks/mimo_v1 at a8acfde13cc1be1672da3d17ee70bc4d842b4c19; weights Qwen/Qwen3-30B-A3B at ad44e777bcd18fa416d9da3bd8f70d33ebb85d39; contract sha256 1da068e415206dbba66e49ec2b946d08a735bfc48a55957844e523b66ee98d86.

In the records, this server's paths, user name and host name are replaced by , , and ; nothing else was changed.