TG Telegram Group & Channel
Data Science by ODS.ai 🦜 | United States America (US)
Create: Update:

🌟 DeepCoder-14B

New code reasoning LLM fine-tuned from DeepSeek-R1-Distill-Qwen-14B using distributed RL with GRPO+ and iterative context lengthening. Trained on ~24K coding problems (TACO-Verified, PrimeIntellect SYNTHETIC-1, LCB v5), it improves Pass@1 on LiveCodeBench v5 to 60.6%, +7.6% over base and on par with OpenAI o3-mini.

- GRPO+: removes KL/entropy loss for stability; adds offline difficulty filtering, DAPO-inspired loss masking, and reward clipping.
- Iterative context scaling: 16K→32K→64K generalization with improved long-context reasoning.

Eval: Strong results on LiveCodeBench, Codeforces, HumanEval+

Open weights🔥

https://huggingface.co/agentica-org/DeepCoder-14B-Preview

@opendatascience

🌟 DeepCoder-14B

New code reasoning LLM fine-tuned from DeepSeek-R1-Distill-Qwen-14B using distributed RL with GRPO+ and iterative context lengthening. Trained on ~24K coding problems (TACO-Verified, PrimeIntellect SYNTHETIC-1, LCB v5), it improves Pass@1 on LiveCodeBench v5 to 60.6%, +7.6% over base and on par with OpenAI o3-mini.

- GRPO+: removes KL/entropy loss for stability; adds offline difficulty filtering, DAPO-inspired loss masking, and reward clipping.
- Iterative context scaling: 16K→32K→64K generalization with improved long-context reasoning.

Eval: Strong results on LiveCodeBench, Codeforces, HumanEval+

Open weights🔥

https://huggingface.co/agentica-org/DeepCoder-14B-Preview

@opendatascience
Please open Telegram to view this post
VIEW IN TELEGRAM


>>Click here to continue<<

Data Science by ODS.ai 🦜






Share with your best friend
VIEW MORE

United States America Popular Telegram Group (US)