ai-and/termgrade-bandcentral-gemma4-31b-bash

TermGrade Band-Central: the 151-task selection and its splits

_{Part of TermGrade: graded environments and trajectories for terminal agents. Read the blog post.}

271 executable environments · every task carries a measured base pass rate · the exact RL split we trained on

The 151 training and 120 validation environments behind ai-and/termgrade-gemma4-31b-lora, with the verl config that produced it.

_{Companion releases: ai-and/termgrade-environments (1,004 environments graded by six models) · ai-and/termgrade-trajectories (36k graded episodes)}

---

Quick start

from datasets import load_dataset

train = load_dataset("ai-and/termgrade-bandcentral-gemma4-31b-bash", "train151")["train"]
print(train[0]["extra_info"]["verl_band_pass_rate"])    # 0.5 for every training task

Build the base image once. Every task Dockerfile is FROM dsbase-venv312:local, and every tests/test.sh calls /opt/verifier_venv/bin/pytest, which lives in the base:

docker build -t dsbase-venv312:local base_image/

Then build a task and run its reference solution against its own verifier, offline:

T=environments/ds2-task_00104_f0fb18
docker build -t mytask:local $T/environment
c=$(docker run -d --network none mytask:local sleep 600)
docker exec $c sh -c 'mkdir -p /tests /logs/verifier'
docker cp $T/tests/.            $c:/tests/
docker cp $T/solution/solve.sh  $c:/solve.sh
docker exec $c sh -c 'cd /app && bash /solve.sh'
docker cp $T/tests/.            $c:/tests/     # restore pristine, as grading does
docker exec $c sh -c 'bash /tests/test.sh'
docker exec $c cat /logs/verifier/reward.txt    # -> 1
docker rm -f $c

The image has no test files. The suite is copied in at scoring time so the agent cannot read what it is graded on. Copy it in yourself, as above, or the tests cannot run.

Some SETA-derived reference solutions (28 of the 59) need network access or packages that may no longer install, so they can fail in an offline rebuild.

---

What ships

Path	Contents
data/train_bandcentral151.parquet	151 tasks, base pass rate 0.50 (118 DeepSeek-generated, 33 SETA)
data/val_heldout120.parquet	120 tasks, disjoint from train: 60 at p≈0.25, 60 at p≈0.75
environments/<seed_id>/	all 271 environments: environment/, instruction.md, tests/, solution/
base_image/Dockerfile	the shared base every task builds FROM
training_config.yaml	the exact verl / GRPO++ configuration

Ids carry a ds2- or setaevolve-/setasynth- prefix by source. Strip the prefix before comparing against ai-and/termgrade-environments, or a plain set comparison finds zero overlap. 113 of the 151 and 89 of the 120 also appear in that 1,004-task corpus, labelled there as train151 / heldout120 for decontamination.

---

The selection rule

For GRPO the gradient contribution of a prompt scales with the variance of its rollout rewards. With binary rewards that variance is p(1−p), maximal at p = 0.50. A task the model always solves and one it never solves both contribute nothing. So the rule is: grade every candidate with the model you intend to train, then take the ones nearest 0.50.

These labels were measured in a bash tool-calling scaffold, which is what we train in. The six-model grades in ai-and/termgrade-environments were measured under Terminus-2, so the two sets of numbers are not directly comparable.

---

Overlap with our released checkpoint

This is training data. 271 of these environments were seen by our released checkpoint. For a clean evaluation use the 802-task unseen split of ai-and/termgrade-environments.

Difficulty depends on the model. Across six models the "solved about half the time" sets barely overlap: pairwise Jaccard 0.11–0.21 (0.29 for gemma-4-31B-it against itself with thinking on), and no task is band-central for all six. These labels are valid for gemma-4-31B-it under a bash scaffold only. If you select training data this way, regrade against your own model.

---

Result

The checkpoint trained on this split scores 46.1 on Terminal-Bench 2.1 against a 43.0 base, so +3.1. Across five independent runs of the identical recipe the mean is 45.1 (+2.1), SE 0.6, all five above the base model. Full methodology on the adapter card.

Citation

If you use TermGrade, please cite:

@misc{calik2026termgrade,
  title        = {{TermGrade}: 1k Graded Terminal Environments, 36k Trajectories, and the {RL} Run They Trained},
  author       = {Calik, Yagiz and Wu, Jianbo and Hara, Shimpei},
  year         = {2026},
  month        = oct,
  howpublished = {\url{https://www.aiand.com/newsroom/termgrade}},
  note         = {ai\& Research blog post}
}