Ripple-65B Preview Achieves U.S. Open-Weight SOTA on Agentic Coding
Original title: Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
The model uses custom synthetic data and RL recipes to reach 80.5 TerminalBench with 1.4M rollouts, 70x fewer than prior SOTA.
Recursive's Ripple-65B preview carefully iterated on SFT and RL recipes with focus on compute efficiency. It opted not to buy posttraining data and devised its own synthetic data pipelines, quality filtering and coverage metrics. The mixed RL recipe uses truncated importance sampling, score centering and linear length penalty with high-quality long-horizon synthetic environments. The RL tasks were tuned for model capabilities and harder tasks were added as the model improved. The end result was a steep RL climb from a strong initial SFT checkpoint.
- 80.5 TerminalBench score with only 1.4M rollouts
- 70x fewer rollouts than Reflection Beam
- No posttraining data purchased
- Custom synthetic data and RL tuning for agentic coding
- AI models
- Ripple-65B
From the original article
1 min read at ricursive.com
Rather than investing in a brute-force RL climb, we carefully iterated on our SFT and RL recipes with a focus on compute efficiency. We opted not to buy any posttraining data, instead devising our own synthetic data pipelines, quality filtering, and coverage metrics. with high-quality long-horizon synthetic environments.
By Philpax · ricursive.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.