SpecBench: Turning Intent into Specifications
SpecBench benchmarks how well AI coding agents collaborate with users to turn vague ideas into structured specifications, and introduces Buddy, an agent that drafts better specs with fewer questions.
Open-source AI
A collective of researchers and engineers from Red Hat & IBM building LLM toolkits you can use today.
LLM Hubs
its_hub
Inference-time scaling for LLMs.
sdg_hub
Synthetic data generation pipelines
training_hub
Post training algorithms for LLMs
LLM Tools
async-grpo
Asynchronous GRPO for scalable reinforcement learning.
hopscotch
A method for skipping redundant attention blocks in language models
mini_trainer
Efficient training library for large language models up to 70B parameters on a single node.
orthogonal-subspace-learning
Adaptive SVD-based continual learning method for LLMs.
probabilistic-inference-scaling
Inference-time scaling with particle filtering.
reward_hub
State-of-the-art reward models for preference data generation and acceptance criteria.
SQuat
KV cache quantization for scaling inference time
training
Efficient messages-format SFT library for language models
Recent posts
View all posts
Unsloth and Training Hub: Lightning-fast LoRA and QLoRA fine-tuning
Training Hub v0.4.0 adds LoRA and QLoRA fine-tuning powered by Unsloth, enabling fast, cost-effective model adaptation with roughly 70% less VRAM than full fine-tuning.
Scale LLM fine-tuning with Training Hub and OpenShift AI
A four-step pathway for scaling LLM fine-tuning from local experimentation to production deployment using Training Hub, OpenShift AI, Kubeflow Trainer, and AI pipelines.
Recent videos
View all videos📹 Towards Self-Adapting Models
👤 Speaker: Jyo Pari
📹 Self-Distillation as a New Framework for Continual Learning
👤 Speaker: Idan Shenfeld
📹 Reasoning with Sampling: Your Base Model is Smarter Than You Think
👤 Speaker: Aayush Karan