Add GEVal logprob artifacts for OpenAI and Fireworks

[benjibc]

Summary

• point the local filesystem dataset logger at the shared directory utils so it can be imported without sqlite
• update the GEval logprob example to run with a filesystem logger and Fireworks/OpenAI completion parameters
• capture JSONL artifacts from running the example with OpenAI and Fireworks models

Testing

• HOME=/workspace/python-sdk/.ep_home_openai EP_COMPLETION_PARAMS='[{"model":"gpt-3.5-turbo","logprobs":true,"top_logprobs":3}]' python -m pytest examples/deepeval/test_geval_with_logprobs.py -k test_geval_with_logprobs -vv --disable-warnings --maxfail=1
• HOME=/workspace/python-sdk/.ep_home_fireworks EP_COMPLETION_PARAMS='[{"model":"accounts/fireworks/models/qwen3-8b","logprobs":true,"api_base":"https://api.fireworks.ai/inference/v1","custom_llm_provider":"fireworks_ai"}]' python -m pytest examples/deepeval/test_geval_with_logprobs.py -k test_geval_with_logprobs -vv --disable-warnings --maxfail=1
• pre-commit run --files eval_protocol/dataset_logger/local_fs_dataset_logger_adapter.py examples/deepeval/test_geval_with_logprobs.py

---

Codex Task

---

Note

Captures provider logprobs on assistant messages during rollouts, updates models and logger import, and adds a GEval example with JSONL artifacts plus tests.

• Evaluation pipeline:
  • Attach provider logprobs to assistant Message during single-turn rollouts (eval_protocol/pytest/default_single_turn_rollout_process.py), with _serialize_logprobs for JSON-safe payloads.
  • Extend Message model with optional logprobs and exclude it from request payloads (eval_protocol/models.py).
• Dataset logger:
  • Point local FS logger to shared directory utils (eval_protocol/dataset_logger/local_fs_dataset_logger_adapter.py).
• Examples & Artifacts:
  • Add GEval example that logs to local FS and forwards final-message logprobs into metric data (examples/deepeval/test_geval_with_logprobs.py).
  • Include JSONL artifacts for OpenAI, Fireworks, and combined runs (examples/deepeval/artifacts/*.jsonl).
• Tests:
  • Add unit test ensuring rollout captures and stores logprobs (tests/test_rollout_logprobs.py).

^{Written by Cursor Bugbot for commit ff210d4. This will update automatically on new commits. Configure here.}