Add GSM-Infinite training environment

[poofeth]

Summary

• add a GSM-Infinite train/eval environment backed by the public InfiniAILab Hugging Face datasets
• support medium/hard/symbolic dataset variants, context-length suffixes, operation splits, and optional train/eval example limits
• normalize GSM-Infinite rows into Verifiers question/answer/info examples and score final Answer: ... values with an exact-answer reward

Bounty: https://algora.io/PrimeIntellect-ai/bounties/m4592Ap5jB1qm9yd

Submitted as the custom Algora bounty claim for GSM-Infinite.

Validation

• uv run pytest tests/test_gsm_infinite_environment.py -q
• uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
• uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
• git diff --check
• live smoke-loaded one row from InfiniAILab/gsm_infinite_medium_0 / ops_2 and verified symbolic dataset naming resolves to InfiniAILab/gsm_infinite_symbolic_example_0

---

Note

Medium Risk
Introduces a new HF-backed environment and answer parsing/reward logic; main risk is dataset naming/split parameters and answer extraction edge cases affecting scoring correctness.

Overview
Adds a new installable gsm-infinite SingleTurn environment backed by InfiniAILab’s Hugging Face GSM-Infinite datasets, with configurable subset/context-length/operation split and optional train/eval example limits.

Implements row normalization into question/answer/info plus an exact-match reward based on the final Answer: ... in the model output, and registers the environment via a new pyproject.toml entry point; includes docs and focused tests for dataset naming, mapping, limiting, and reward parsing.

^{Reviewed by Cursor Bugbot for commit 76ab25b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[poofeth]

Validation evidence for the bounty review:

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

$ uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
All checks passed!

$ uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
2 files already formatted

$ git diff --check
# no output

$ uv run python - <<'PY'
# loaded one example through load_gsm_infinite_dataset(num_examples=1)
# result: 1 ['question', 'answer', 'info']; answer: 2
# symbolic name check: InfiniAILab/gsm_infinite_symbolic_example_0
PY

This PR targets https://algora.io/PrimeIntellect-ai/bounties/m4592Ap5jB1qm9yd.

[poofeth]

Proactive follow-up in commit a9d549a:

• added the gsm-infinite environment entry point
• included README.md in the package build
• listed gsm_infinite in environments/README.md

Validation after the update:

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

$ uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
All checks passed!

$ uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
2 files already formatted

$ git diff --check
# no output

[poofeth]

Addressed the Bugbot answer-extraction feedback in commit 203cfb1:

• collapsed the unreachable duplicate Answer: regex path into one case-insensitive prefix parser
• avoid re-extracting expected answers during reward normalization, preserving multi-value list answers
• added regression coverage for multi-value answers and prefixed numeric/symbolic extraction

Validation:

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

$ uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
All checks passed!

$ uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
2 files already formatted

$ git diff --check
# no output

[poofeth]

Addressed the latest Bugbot follow-up in commit 4f732d8:

• extract_answer now uses the final Answer: marker instead of the first marker, matching the prompt instruction to end with the final result
• eval_subset, eval_context_length, and eval_operations now only fall back on None, preserving explicit falsy values
• added regression coverage for intermediate/final answer markers and eval_operations=0

Validation:

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

$ uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
All checks passed!

$ uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
2 files already formatted

$ git diff --check
# no output

[poofeth]

Addressed the latest Bugbot symbolic dataset finding in commit e1e9a26:

• changed symbolic dataset resolution from InfiniAILab/gsm_infinite_symbolic_example_{context_length} to the full dataset names InfiniAILab/gsm_infinite_symbolic_{context_length}
• added regression coverage for both symbolic, context_length=0 and symbolic, context_length=8k
• verified the Hugging Face dataset API currently resolves InfiniAILab/gsm_infinite_symbolic_0, InfiniAILab/gsm_infinite_symbolic_8k, and InfiniAILab/gsm_infinite_symbolic_16k

Validation after the fix:

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

$ uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
All checks passed!

$ uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
2 files already formatted

$ git diff --check
# no output

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit e1e9a26. Configure here.}

[poofeth]

Addressed the latest Bugbot answer_list=None fallback finding in commit 76ab25b:

• row_to_example now falls back to solution when the answer_list key exists but its value is None or otherwise empty
• added regression coverage so a row with answer_list = None extracts the expected answer from solution instead of producing "None"

Validation:

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

$ uv run ruff check environments/gsm_infinite tests/test_gsm_infinite_environment.py
All checks passed!

$ uv run ruff format --check environments/gsm_infinite tests/test_gsm_infinite_environment.py
2 files already formatted

$ git diff --check
# no output

[poofeth]

/claim https://algora.io/PrimeIntellect-ai/bounties/m4592Ap5jB1qm9yd