feat(gsm8k): add dataset and 8-shot base-model task

[jack-scitix-ai]

Type

• feature — new benchmark, task, or capability

Summary

• Add the GSM8K dataset wrapper backed by hf:openai/gsm8k.
• Add gsm8k_8shot_base_gen, an 8-shot generative task for base models.
• Mirror lm-evaluation-harness GSM8K prompt, stop sequences, and answer extraction behavior, with num_fewshot set to 8.
• Register the GSM8K dataset/task exports and meta index entries for discovery.
• Reference paper: https://arxiv.org/abs/2110.14168
• Reference repo: https://github.com/EleutherAI/lm-evaluation-harness/blob/1dd931087362abba74e0375c8c631295559f48b2/lm_eval/tasks/gsm8k/gsm8k.yaml

Test Plan

Automated

• Lint/format clean (ruff check && ruff format --check)
• Type check clean (ty check or mypy --strict)

Manual

• Dataset loading tested (pdm run sieval dataset download gsm8k --data-dir /tmp/sieval-gsm8k-download-test) succeeds
• Ran gsm8k_8shot_base_gen on the full GSM8K test split with Qwen2.5-72B Base
• Run completed 1319/1319 samples with 0 failures
• Actual score: 86.35 exact match; reference score: 88.30; diff: -1.95

Model	Expected	Actual	Diff
Qwen2.5-72B Base	88.30	86.35	-1.95(-2.21%)

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring
• No new upper-layer dependencies added to core/
• Deleted code verified — no remaining call sites depend on it

If: New or Modified Benchmark

• Reference paper/repo linked in Summary
• Score comparison table included (model, expected, actual, diff)
• Dataset loading tested (sieval dataset download <name> succeeds)
• Task registered in package-level __init__.py