GPT-4 pretrained base (paper model)

ClosedAnnounced, no public access

The original GPT-4 report identifies a pretrained base model for its traditional benchmark evaluations, separate from subsequent alignment. This entry pins arXiv v1, submitted March 15, 2023. It is not an identified ChatGPT or API checkpoint; its exact architecture and service availability are unspecified.

Benchmark scores

HellaSwag · Reported by the model's maker, Mar 15, 202395.3%#1 of 5 results
HumanEval · Reported by the model's maker, Mar 15, 202367.0%#2 of 4 results
MMLU · Reported by the model's maker, Mar 15, 202386.4%#1 of 2 results

Access

Research benchmark target described in the original v1 report, before RLHF. No served API identifier or public/API/weights-access date is established for this exact pretrained checkpoint. Text capability is documented by its next-token task and traditional language benchmarks; other exact base-model specifications remain unspecified.