Model rollouts, scored one completion at a time

Millions of generations across code and structured-generation domains, each individually scored by a deterministic evaluator and browsable down to the single completion with its prompt, verdict, and error. Every curve carries a 95% bootstrap confidence interval.

Contents

DS-1000

The 1000-problem sweep in full: temperature curves, significance tests, and the complete error analysis, including Qwen2.5-Coder and DeepSeek-Coder.

LiveCodeBench

Competitive programming (714 problems, 2023-05 to 2024-09): temperature curves with CIs, failure categories, and bug themes. 6 models complete; more being added.

Other domains

The same temperature sweep on three non-code domains (molecular synthesis, goal inference, text-to-SQL) shown next to DS-1000, for contrast with the code benchmarks above.

In preparation

Constrained decoding (SMC)

The same tasks decoded under sequential Monte Carlo with grammar constraints, compared against raw sampling.

LiveCodeBench

The sweep methodology applied to competitive-programming problems.

The experiment

For each domain, each model draws 100 completions per instance at every temperature with plain vLLM sampling (no constrained decoding, no SMC). Every completion is scored individually; pass@k uses the unbiased estimator, averaged over instances, with confidence intervals from 2000 bootstrap resamples of the instance set and paired tests for peak temperatures and temperature steps.

DomainTaskInstancesScored by
DS-1000data-science code completion, 7 libraries1000 official test execution + source check, fresh subprocess, deterministic
Molecular synthesisgenerate drug-like molecules (SMILES)100 RDKit validity + QED
Goal inferenceinfer PDDL goals (blocksworld)100 planetarium goal equivalence
Text-to-SQLSpider100 SQLite execution match, sequential scoring

The data

Everything lives in a public Hugging Face dataset, samuki-hf/temperature-sweep-data: hive-partitioned Parquet, one partition per (domain, model, temperature). Each generation carries its verdict (passed, plus score/valid where applicable) and joins to the prompt, failure, and bug-theme tables by (instance_id, sample).

DuckDB queries it in place, so a question costs seconds, not a download, and the dataset is public: no token needed. The repo's CLAUDE.md documents every schema, so you can also clone the repo, run Claude Code, and ask questions about the data directly.

import duckdb
con = duckdb.connect()
con.sql("""
  SELECT model, temp, avg(passed::INT) AS pass_rate
  FROM read_parquet(
    'hf://datasets/samuki-hf/temperature-sweep-data'
    '/rollouts/**/*.parquet', hive_partitioning=1)
  WHERE domain = 'ds1000' AND passed IS NOT NULL
  GROUP BY model, temp ORDER BY model, temp
""").df()