T2J-Bench: leaderboard
Codebase conversion benchmark (T2J-Bench; Converted, Not Equivalent, Song et al., 2026): the agent converts 45 LLaMA-Factory-style PyTorch training codebases (15 model families, each with SFT, DPO and PPO trainers) to JAX, running inside the authors' shared multi-agent coding scaffold. A fixed verifier checks the interface, numeric parity of outputs, losses and gradients, and short training behaviour under fixed seeds; the score is the percent of codebases whose first conversion passes all three stages, higher is better.
Metric: Conversion Pass Rate (%; first attempt passes all three verifier stages, 45 codebases). Source: arxiv.org. Saturation forecast: Around April 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 28.9 | #75 |
| 2 | Claude Sonnet 4.6 | 22.2 | #158 |
| 3 | Gemini 3.1 Pro (Preview) | 13.3 | #91 |
| 4 | Claude Opus 4.6 | 11.1 | #100 |
| 5 | GLM-5 | 6.7 | #256 |
| 6 | Qwen 3 Coder 480B A35B Instruct | 0 | #641 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=t2j-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.