Add IFBench eval suite (ifeval_ood + ifeval_mt) aligned with oe-eval-internal

[finbarrtimbers]

Adds the IFBench instruction-following benchmark to olmo-eval-internal, matching the ifbench::tulu suite from oe-eval-internal.

Validation

uv run olmo-eval beaker launch -y \
    -m Qwen/Qwen3-4B \
    --harness default \
    -B ai2/oe-adapt \
    -o provider.kind=vllm_server \
    -t ifbench \
    --gpus 1 -c h100 -w ai2/open-instruct-dev -p urgent

Reference (allenai/oe-eval-internal):

uv run python oe_eval/launch.py \
    --model Qwen/Qwen3-4B \
    --model-type vllm \
    --model-args '{"trust_remote_code": true, "max_length": 32768}' \
    --task ifbench::tulu \
    --gpus 1 --cluster ai2/jupiter \
    --beaker-workspace ai2/open-instruct-dev \
    --beaker-priority urgent --beaker-budget ai2/oe-adapt

Results (Qwen3-4B, prompt_level_loose_acc)

sub-task	n	oe-eval-internal	this PR
ifeval_mt_wildchat_unused_withRewrite	1774	0.527	0.544
ifeval_mt_ood_wildchat_unused_withRewrite	1387	0.358	0.360
ifeval_ood	300	0.213	0.240
ifbench (suite average)	—	—	0.381

Beaker run: https://beaker.org/ex/01KQTMJNA0YMFJ3AVSSAVG4G1A

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1e850b7565

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[undfined]

looks good!