AI research and Engineering
for Human Intelligence

TIMPS-Coder-7B — 98.8% HumanEval, 4-stage pipeline

A 7B-parameter code model fine-tuned via SFT, GRPO, DPO, and Self-Play. MBPP+ 73.3% — the clean, uncontaminated benchmark. HumanEval results disclosed with full transparency on training-data overlap. Apache-2.0.