Add VLLMSoftEntailer for LLM-based conditional probability estimation#2
Add VLLMSoftEntailer for LLM-based conditional probability estimation#2
Conversation
Implement a new entailer class that uses a VLLM-backed OpenAI-compatible server endpoint hosting Zhengping/conditional-probability-regression. The entailer estimates p(h|p) by extracting the first-token distribution over special <|label_level_N|> tokens and computing a softmax-weighted average of their midpoint scores, producing a probability in [0, 1]. https://claude.ai/code/session_018eo6tgjgbqwGcoaaf45K2L
Replace urllib-based HTTP handling with the official openai Python client for cleaner, typed interaction with the vLLM OpenAI-compatible server. Add openai as a project dependency in pyproject.toml and requirements.txt. https://claude.ai/code/session_018eo6tgjgbqwGcoaaf45K2L
Replace sequential sync OpenAI calls with concurrent async requests using AsyncOpenAI and asyncio.gather. The client is created per-batch inside an async context manager to avoid stale event-loop bindings. Built-in exponential-backoff retries (configurable via max_retries) handle transient failures (connection errors, HTTP 429/5xx). A per-request timeout is also configurable. https://claude.ai/code/session_018eo6tgjgbqwGcoaaf45K2L
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a04cb033e2
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".
| errors (connection failures, HTTP 429 / >=500) with | ||
| exponential backoff. | ||
| """ | ||
| return asyncio.run(self._async_call_batch(instances)) |
There was a problem hiding this comment.
_call_batch unconditionally wraps every batch in asyncio.run, which raises RuntimeError whenever the caller already has an active event loop (e.g., Jupyter notebooks, pytest-asyncio, FastAPI workers). In those common environments this new entailer cannot be used at all, so experiments that switch to VLLMSoftEntailer will fail before scoring; this path is reached through the normal Entailer.__call__ flow, not just a special API.
Useful? React with 👍 / 👎.
2 participants
Implement a new entailer class that uses a VLLM-backed OpenAI-compatible
server endpoint hosting Zhengping/conditional-probability-regression.
The entailer estimates p(h|p) by extracting the first-token distribution
over special <|label_level_N|> tokens and computing a softmax-weighted
average of their midpoint scores, producing a probability in [0, 1].
https://claude.ai/code/session_018eo6tgjgbqwGcoaaf45K2L