[rsareddy0329]

What

Adds the required prompt field to each record in
sagemaker-mlops/tests/integ/data/v3_trainer/preference_train.jsonl, converting it to the
documented DPO schema prompt + chosen + rejected (prompt = the user turn;
chosen/rejected = the assistant completions).

Why

test_dpo_trainer_pipeline_create_and_execute (tests/integ/workflow/test_v3_trainer_pipeline.py)
started failing on the 2026-10-29 CI Health V3 Master canary:

DPOFineTune: ClientError: Dataset is missing required columns: 'prompt' (set via data.prompt_key).
Available columns: ['chosen', 'rejected']

Root cause is external JumpStart Hub drift, not an SDK change:

Event	Time (UTC)
DPO recipe v2.0.6 published to JumpStart cache	2026-10-28 05:29
Hub content 2.17.0 for meta-textgeneration-llama-3-2-1b-instruct goes live	2026-10-28 12:37
CI Health canary, DPO PASSED (prior hub content)	2026-10-28 07:04
CI Health canary, DPO FAILED (now on 2.17.0)	2026-10-29 07:04

The republished training container now enforces the prompt column the recipe maps via
data.prompt_key. The fixture only carried chosen/rejected. The only SDK commit merged in that
window was an unrelated codegen/codestyle fix, so no SDK code caused this.

Not a customer-facing change

prompt + chosen + rejected is the format the DPO example notebook documents and the schema the
SDK's own sagemaker-train/.../ai_registry/dataset_validation.py::validate_dpo enforces (it requires
all three). This only realigns a stale test fixture with the already-documented, already-validated
contract.

Validation

• DPO integ run to confirm the container accepts the conversational shape (prompt = user turn,
  chosen/rejected = assistant completions). The exact shape is defined by the verl DPODataConfig
  in the 2.17.0 container; can switch to plain strings if the container expects that instead.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[github-actions]

🤖 AI code review

This PR updates a single integ-test fixture (preference_train.jsonl) to add a prompt field to each DPO record, splitting the previous chosen/rejected conversational turns into prompt (user turn) + chosen/rejected (assistant completions). The fixture is uploaded verbatim to S3 and consumed by the training container in test_dpo_trainer_pipeline_create_and_execute; it does not pass through the SDK's validate_dpo, so the change is self-consistent. The edit is mechanical and consistent across all 20 rows, and it realigns the data with the container's reported required columns. The change looks correct; I have no blocking findings, only noting that the PR author's own checklist flags that the conversational list shape (vs. plain strings) still needs confirmation via an actual DPO integ run.

_{Reviewed commit fc55730. Automated review; verify before acting.}