Add Spiral-Bench synthetic training environment

[poofeth]

Summary

• add a spiral_bench environment inspired by Spiral-Bench safety conversations
• include a deterministic prompt generator for uncontaminated synthetic training prompts
• commit a 64-row HF Dataset-compatible JSONL sample with question, answer, and info fields
• add focused tests covering generator determinism, sample dataset loading, and environment construction without API calls

Bounty

• Algora: https://algora.io/PrimeIntellect-ai/bounties/gdFDkFi3PzcwFndN
• Claiming the Spiral-Bench + train set bounty.

Validation

• uv run pytest tests/test_spiral_bench_environment.py -q
• uv run ruff check environments/spiral_bench tests/test_spiral_bench_environment.py
• uv run ruff format --check environments/spiral_bench tests/test_spiral_bench_environment.py
• git diff --check

---

Note

Low Risk
Low risk: the PR primarily adds a new standalone environment package, sample data, and tests, without changing shared runtime logic; main risk is limited to correctness of the new judge-based reward parsing/scaling.

Overview
Adds a new spiral_bench single-turn environment with a judge-scored 0–10 safety rubric, including dataset loading (build_dataset) and a normalized reward function wired into vf.SingleTurnEnv.

Includes a deterministic synthetic prompt generator and commits a 64-row JSONL sample dataset, plus docs and packaging (pyproject.toml) so the environment can be installed and regenerated. Updates environments/README.md to list spiral_bench, and adds targeted tests covering generator determinism/limits, dataset loading, and that load_environment() constructs without requiring an API key.

^{Reviewed by Cursor Bugbot for commit 0b9b606. Bugbot is set up for automated code reviews on this repo. Configure here.}

[poofeth]

Validation evidence for the Spiral-Bench bounty PR:

• uv run pytest tests/test_spiral_bench_environment.py -q passed: 3 tests.
• uv run ruff check environments/spiral_bench tests/test_spiral_bench_environment.py passed.
• uv run ruff format --check environments/spiral_bench tests/test_spiral_bench_environment.py passed.
• git diff --check passed.

The sample dataset is generated from local deterministic templates (source=synthetic-uncontaminated-v1) rather than copied from the upstream Spiral-Bench benchmark prompts.

[poofeth]

Addressed the Bugbot review in commit 16a0b72:

• fixed the judge call so JudgeRubric receives the original prompt/completion and sees the actual assistant response
• added an upper-bound guard for synthetic generation (num_examples <= 600) to avoid duplicate-exhaustion loops
• added spiral_bench to environments/README.md
• added regression coverage for the generator bound

Validation after the fixes:

$ uv run pytest tests/test_spiral_bench_environment.py -q
....                                                                     [100%]

$ uv run ruff check environments/spiral_bench tests/test_spiral_bench_environment.py
All checks passed!

$ uv run ruff format --check environments/spiral_bench tests/test_spiral_bench_environment.py
3 files already formatted

$ git diff --check
# no output

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 16a0b72. Configure here.}

[poofeth]

Addressed the latest Bugbot follow-up in commit 0b9b606:

• removed dead _completion_text helper after switching to JudgeRubric native prompt/completion handling
• added README.md and pyproject.toml to Hatch build includes
• listed spiral_bench under the judge-based evaluation section in environments/README.md

Validation:

$ uv run pytest tests/test_spiral_bench_environment.py -q
....                                                                     [100%]

$ uv run ruff check environments/spiral_bench tests/test_spiral_bench_environment.py
All checks passed!

$ uv run ruff format --check environments/spiral_bench tests/test_spiral_bench_environment.py
3 files already formatted

$ git diff --check
# no output

[poofeth]

/claim https://algora.io/PrimeIntellect-ai/bounties/gdFDkFi3PzcwFndN