Made reinforcement learning from human feedback (RLHF) the standard alignment recipe: collect demonstrations and preference rankings, train a reward model, then optimize with PPO. A 1.3B aligned model was preferred over the 175B GPT-3 by human raters.
AI Dataset2022
Benchmark dataset of ~8.5k grade-school math word problems with step-by-step solutions and calculator annotations for evaluating multi-step arithmetic reasoning in language models. Provided in two configs (main and socratic) and commonly used for chain-of-thought prompting, fine-tuning, and verifier training.
A line-by-line PyTorch reimplementation of the Transformer paper as a runnable notebook, where each part of the paper sits next to the code that implements it — turning a dense architecture into something you can read and run end to end.
AI Model2022
Rust-and-Python toolkit that serves open-source LLMs (Llama, Falcon, Mixtral, StarCoder) over HTTP/gRPC with tensor parallelism, continuous batching, Flash/Paged Attention and quantization. Now in maintenance mode, pointing users toward vLLM and SGLang.
AI Dataset2022
Provides human preference comparison pairs and red-team conversation transcripts collected by Anthropic for training preference/reward models and studying harmful model behaviors; intended for RLHF and safety research, not for supervised fine-tuning of dialogue agents.
AI Dataset2022
Community-curated collection of ChatGPT-style prompts mirrored as a Hugging Face dataset; organized by task and model compatibility for quick reuse. Useful for prompt engineering, text-generation prototyping, and building conversational examples across multiple LLMs.
AI Train2022
Reproduces GPT-2 (124M) from scratch on OpenWebText in ~4 days on an 8xA100 node, with the whole stack kept to two ~300-line files: train.py for the loop and model.py for the architecture. A char-level Shakespeare run finishes in ~3 minutes on one GPU.
AI Infra2023
Lets you write compositional Python programs that compile into self‑improving LLM pipelines — replacing brittle prompt engineering with a declarative, programmatic approach and built‑in algorithms to optimize prompts and weights for RAG, multi‑stage pipelines, and agent loops.
AI Agent2023
Runs an agentic RAG loop over scientific papers: searches literature, gathers and re-ranks evidence chunks, then answers with in-text citations. Adds metadata-aware embeddings, retraction checks, and contradiction detection across full PDFs.
AI Dataset2023
Contains tech-blog posts scraped from Habr (primarily Russian, some English) in Parquet format with ~100K–1M records. Suited for multilingual text-generation and language-model fine-tuning; license is not specified, so verify before redistribution.
AI Dataset2023
Provides 52,000 English instruction–response pairs generated by OpenAI's text-davinci-003 for instruction-tuning language models. Released under CC BY-NC 4.0; low-cost synthetic data useful for research but contains model-generated biases and errors.
A multimodal model that accepts image and text inputs and returns text, scoring at human level on professional exams — including a bar exam in the top 10%. Its performance was forecast from models using 1/1000th the compute, showing predictable scaling.