[PHASE IV] fix(llm_simple_qa): pass do_sample explicitly so the benchmark is reproducible

[vjuhhii]

What type of PR is this?

/kind bug

What this PR does / why we need it:

examples/llm_simple_qa scores differently on every run. Four end-to-end runs of unchanged code and
configuration produced 0.6, 0.7, 0.8 and 0.9, reported in #988.

The cause is one missing argument. basemodel.py calls

generated_ids = self.model.generate(
    model_inputs.input_ids,
    max_new_tokens=512
)

Because do_sample is not passed, its value comes from the model's own generation_config.json.
For Qwen/Qwen2.5-0.5B-Instruct that file sets do_sample=True, temperature=0.7, top_p=0.8, top_k=20. The example never asked for stochastic decoding; it inherits chat defaults from a
third-party model card that can also change without any commit here.

This passes do_sample=False explicitly. Greedy decoding is the usual choice for evaluation, and it
removes the dependency on an external file for a property the benchmark depends on.

Verification

On a GTX 1080, workspace/benchmarkingjob/rank/all_rank.csv after four consecutive runs of the
unchanged job:

Before (from #988):

0.9   2026-08-28 20:31:40
0.8   2026-08-28 20:41:33
0.7   2026-08-28 20:34:55
0.6   2026-08-28 20:25:06

After, workspace cleared first, four runs:

0.6   2026-08-28 20:52:55
0.6   2026-08-28 20:55:01
0.6   2026-08-28 20:57:07
0.6   2026-08-28 21:00:11

Four distinct run ids and timestamps, one score. transformers emits no warning about the now-unused
temperature and top_p from the generation config.

Scope

Only llm_simple_qa is changed, deliberately.

Nine other generate() calls exist under examples/. The closest relatives —
government/singletask_learning_bench (objective and subjective) and
smart_coding_learning_bench (comment and issue) — pass temperature=0.1, top_p=0.9 explicitly.
That is an author expressing intent to sample, and silently converting it to greedy decoding would
be a different change requiring a different conversation. They also load models from hardcoded local
paths such as /home/icyfeather/models/, so they cannot run for anyone else today; determinism is
not their first problem.

llm_simple_qa is the only one that passes no decoding arguments at all and therefore inherits the
behaviour silently.

Not addressed here

#988 also raises three related points that are deliberately left out of this PR:

• A configurable seed in testenv.yaml for examples that genuinely intend to sample. That changes
  the config schema and is worth designing with maintainers.
• Recording in the report whether decoding was deterministic. That touches the results pipeline.
• The ten-row test set giving a metric resolution of 0.1, which is a separate discussion.

Which issue(s) this PR fixes:

Fixes #988

[kubeedge-bot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: vjuhhii
To complete the pull request process, please assign jaypume after the PR has been reviewed.
You can assign the PR to them by writing /assign @jaypume in a comment when ready.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment