# PostTrainBench ![Category](https://img.shields.io/badge/Category-ML%20Research%20Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Success-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) [![Website](https://img.shields.io/badge/Website-posttrainbench.com-0969da)](https://posttrainbench.com) > **Quick answer:** PostTrainBench is an agentic benchmark testing AI models on post-training and ML research engineering workflows — tasks like curating fine-tuning data, configuring training runs, and building evaluation pipelines. It targets the emerging use case of AI agents assisting with ML research itself. Kimi K3 scores 36.6% as of July 2026. ## At a Glance **What it tests:** A model's ability to act as an ML research/engineering agent, handling tasks across the post-training lifecycle — data curation, fine-tuning setup, and evaluation. **Why it matters:** As labs increasingly use AI agents to accelerate their own research and post-training pipelines, PostTrainBench measures a distinct and practically important capability: can a model help build the next model? **Known limitations:** Low scores across the field (Kimi K3's 36.6% is a frontier-level result) indicate this remains an early, unsaturated, and difficult benchmark with limited public documentation of exact task composition. ## What PostTrainBench Measures PostTrainBench evaluates AI agents on realistic ML post-training research and engineering tasks, such as preparing and filtering fine-tuning datasets, configuring RLHF/RLVR training pipelines, and constructing evaluation harnesses. It reflects a meta-capability increasingly relevant to frontier labs: using AI to help accelerate the development of future AI systems. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent / ML research engineering | | Metric | % task success | | Saturation | Low | | Created by | PostTrainBench project | | Website | [posttrainbench.com](https://posttrainbench.com) | ## How PostTrainBench Is Scored Agents attempt ML research/engineering tasks (e.g., data pipeline construction, training configuration) and are scored on % task success, verified against reference implementations or expected pipeline outputs. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Kimi K3](/models/moonshot-ai/kimi-k3) | 36.6% | [Kimi K3 technical report](https://www.kimi.com/blog/kimi-k3) | 2026-07 | *Score sourced from Moonshot AI's Kimi K3 announcement, July 2026.* ## PostTrainBench on Benchgen > No Benchgen results yet — [be the first to run PostTrainBench](/benchmarks/ellis-institute-tubingen/posttrainbench). ## PostTrainBench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **PostTrainBench** | ML post-training research/engineering agent tasks | Low | | [MLS-Bench-Lite](/benchmarks/lyu-et-al/mls-bench-lite) | ML systems engineering tasks | Low | | [SWE-Marathon](/benchmarks/abundant-ai/swe-marathon) | Long-horizon software engineering | Low | | [DeepSWE](/benchmarks/datacurve/deepswe) | Real-world software engineering tasks | Low | ## Run PostTrainBench on Your Model Benchgen lets you run PostTrainBench against your own model, tracking ML research agent capability over time as your team explores using AI to accelerate model development. ## Frequently Asked Questions

What is PostTrainBench?

PostTrainBench is a benchmark testing AI agents on post-training and ML research engineering workflows, including data curation, fine-tuning setup, and evaluation pipeline construction.

What does a good score look like on PostTrainBench?

Kimi K3 reports 36.6% as of July 2026. Given the difficulty of genuine ML research engineering tasks, scores in this range represent frontier-level performance.

Who created PostTrainBench?

PostTrainBench is maintained as an independent benchmark project (posttrainbench.com) focused on evaluating AI agents on ML post-training research workflows.

Rank	Model	Score
1	gemini-4-argon	45.3
2	minimax-m3-1-flash	39.3
3	kimi-k3	36.6
4	kimi-k3-1-pro	36.5

PostTrainBench — Results

PostTrainBench — Phases

PostTrainBench