Deep Dive: Teaching Arcee Trinity Mini to Read Medical Research with RLVR and GRPO

Bojan Jakimovski, an ML engineer, took Arcee AI's open-source Trinity Mini model and turned it into a biomedical specialist — extracting drug-protein relationships from scientific papers. No massive team. No million-dollar budget. Just open weights, a clever training technique called RLVR, and a weekend of GPU time. In this video, I break down exactly how it works: the Mixture of Experts architecture behind Trinity Mini, why Reinforcement Learning with Verifiable Rewards (RLVR) beats traditional fine-tuning for domain specialization, how the GRPO algorithm (the same one behind DeepSeek R1) trains a model to reason step by step, and how LoRA makes it possible to specialize a 26B-parameter model for under $50. Whether you're an ML engineer, a researcher, or just curious about where open-source AI is headed, this is a practical, no-hype walkthrough of a pattern you can replicate in your own domain. Bojan Jakimovski's blog → https://shekswess.github.io Bojan's LinkedIn → https://linkedin.com/in/bojan-jakimovski *** MODELS Trinity-Mini-DrugProt-Think (LoRA adapter) → https://huggingface.co/lokahq/Trinity-Mini-DrugProt-Think Arcee Trinity Mini (base model) → https://huggingface.co/arcee-ai/Trinity-Mini Arcee Trinity Mini Base (pre-SFT) → https://huggingface.co/arcee-ai/Trinity-Mini-Base Trinity Mini on OpenRouter (free tier) → https://openrouter.ai/arcee-ai/trinity-mini:free Trinity Mini on OpenRouter (paid API) → https://openrouter.ai/arcee-ai/trinity-mini *** CODE & CONFIGS Full training repo (configs, metrics, deployment) → https://github.com/LokaHQ/Trinity-Mini-DrugProt-Think 12 experiment TOML configs → https://github.com/LokaHQ/Trinity-Mini-DrugProt-Think/tree/main/experiments/configs/rl Training metrics CSVs → https://github.com/LokaHQ/Trinity-Mini-DrugProt-Think/tree/main/data Deploying on Amazon SageMaker (Loka blog) → https://medium.com/loka-engineering/deploying-trinity-mini-drugprot-think-on-amazon-sagemaker-ai-9e1c1c430ce9 ***DATASETS DrugProt on Hugging Face (bigbio) → https://huggingface.co/datasets/bigbio/drugprot DrugProt Parquet (OpenMed) → https://huggingface.co/datasets/OpenMed/drugprot-parquet *** TOOLS & LIBRARIES Hugging Face Transformers → https://github.com/huggingface/transformers PEFT (LoRA & adapters) → https://github.com/huggingface/peft TRL (GRPOTrainer) → https://github.com/huggingface/trl Prime Intellect (hosted GRPO training) → https://primeintellect.ai Prime Intellect Verifiers (RL environments) → https://github.com/PrimeIntellect-ai/verifiers