Repository navigation
BIRD: output cap, reasoning effort, candidate voting settings and LLM call pacing - #32
Merged
Merged
Conversation
LLM_MAX_OUTPUT_TOKENS caps each reply (sent as max_tokens, the name every OpenAI-compatible API accepts). Cerebras reserves the cap against its rate limits before running a request, so an unset cap reserves the model maximum. LLM_REASONING_EFFORT (low|medium|high) is passed to reasoning models such as gpt-oss. Both are sent only when set, so existing setups are unchanged. Evaluation runs record the LLM host and reasoning effort, the report shows them, a resumed run must keep the same effort, and token estimates count only earlier runs with the same effort. .env.example documents Cerebras.
In the partial Mini-Dev run, 4 of 87 questions were missed only because the model joined two stored columns into one (a definition named them as one term). The rule is general: concatenating is reformatting stored values, which the prompt already discourages. Prompt version sql-generator/7.
- SQL_CANDIDATES (API) and --candidates (evaluation runs) pass the number of candidate queries to AgentController. Runs record it, resume refuses a mismatch, the token estimate only counts runs with the same number, and the report shows it. - LLM_REQUESTS_PER_MINUTE spaces LLM calls (retries included) so a small per-minute limit such as the Cerebras trial's is not hit; with it set the runner's --delay defaults to 0. Pacing waits are not counted in call duration. - Evaluation runs wait out per-minute 429s inside the LLM call (up to 120 s), keeping the question's earlier calls and candidates instead of re-asking. - A 429 asking to wait more than ten minutes counts as a used-up quota even without a body (Cerebras can answer with only Retry-After).
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
First of a stack of six PRs with the BIRD Mini-Dev work. Merge in order.
LLM_MAX_OUTPUT_TOKENS(sent asmax_tokens) andLLM_REASONING_EFFORT(low|medium|high), sent only when set. Evaluation runs record the LLM host and effort, the report shows them, and a resumed run must keep the same effort..env.exampledocuments Cerebras.SQL_CANDIDATES(API) and--candidates(runner) pass the candidate count toAgentController. Runs record it and resume refuses a mismatch.LLM_REQUESTS_PER_MINUTEspaces LLM calls, retries included. Evaluation runs wait out per-minute 429s inside the call (up to 120 s), and a 429 asking to wait more than ten minutes counts as a used-up quota.Testing
Run locally on the top of the stack (
readme-bird-results):ruff check,ruff format --check,mypyclean; backend 340 passed, scripts + evaluation 115 passed (database integration tests skipped locally, run in CI); frontend lint and 57 tests pass. CI runs on each PR.