Ian Maurer
imaurer
·
AI & ML interests
Chief Technical Officer for GenomOncology.
Recent Activity
reacted
to
Ihor's
post
with 🔥
4 months ago
🚀 Reproducing DeepSeek R1 for Text-to-Graph Extraction
I’ve been working on replicating DeepSeek R1, focusing on zero-shot text-to-graph extraction—a challenging task where LMs extract entities and relations from text based on predefined types.
🧠Key Insight:
Language models struggle when constrained by entity/relation types. Supervised training alone isn’t enough, but reinforcement learning (RL), specifically Guided Reward Policy Optimization (GRPO), shows promise.
💡 Why GRPO?
It trains the model to generate structured graphs, optimizing multiple reward functions (format, JSON validity, and extraction accuracy).
It allows the model to learn from both positive and hard negative examples dynamically.
RL can be fine-tuned to emphasize relation extraction improvements.
📊 Early Results:
Even with limited training, F1 scores consistently improved, and we saw clear benefits from RL-based optimization. More training = better performance!
🔬 Next Steps:
We’re scaling up experiments with larger models and high-quality data. Stay tuned for updates! Meanwhile, check out one of our experimental models here:
https://huggingface.co/Ihor/Text2Graph-R1-Qwen2.5-0.5b
📔 Learn more details from the blog post: https://medium.com/p/d8b648d9f419
Feel free to share your thoughts and ask questions!
reacted
to
Ihor's
post
with 🚀
4 months ago
🚀 Reproducing DeepSeek R1 for Text-to-Graph Extraction
I’ve been working on replicating DeepSeek R1, focusing on zero-shot text-to-graph extraction—a challenging task where LMs extract entities and relations from text based on predefined types.
🧠Key Insight:
Language models struggle when constrained by entity/relation types. Supervised training alone isn’t enough, but reinforcement learning (RL), specifically Guided Reward Policy Optimization (GRPO), shows promise.
💡 Why GRPO?
It trains the model to generate structured graphs, optimizing multiple reward functions (format, JSON validity, and extraction accuracy).
It allows the model to learn from both positive and hard negative examples dynamically.
RL can be fine-tuned to emphasize relation extraction improvements.
📊 Early Results:
Even with limited training, F1 scores consistently improved, and we saw clear benefits from RL-based optimization. More training = better performance!
🔬 Next Steps:
We’re scaling up experiments with larger models and high-quality data. Stay tuned for updates! Meanwhile, check out one of our experimental models here:
https://huggingface.co/Ihor/Text2Graph-R1-Qwen2.5-0.5b
📔 Learn more details from the blog post: https://medium.com/p/d8b648d9f419
Feel free to share your thoughts and ask questions!
Organizations
