Provides anonymized multi-domain user behavior sequences and content metadata (short video, ads, e-commerce, live) for cross-domain recommendation, semantic-ID mapping, and content-understanding tasks. Key tables include per-user multi-domain behavior (~500k rows), pid→three-segment semantic IDs, captions, and level-3 tags; all item IDs are hashed for privacy.
A code-agent model for Lean 4 that automates repository-level formal proofs and verification; a Mixture-of-Experts architecture (119B total, 6.5B active) with 256k context, multimodal input and an Apache-2.0 license.
Streams routed experts from disk and treats storage, RAM and VRAM as a single inference hierarchy so MoE models (hundreds of billions to trillions of parameters) can run on consumer or heterogeneous hardware; single-file C engines with zero runtime deps and optional Vulkan/CUDA/Metal backends, speed governed by placement and disk I/O.
Predicts per-request MoE expert footprints from prefill activations and routes decode requests to workers that maximize expert-locality, lowering decode latency by combining offline K-means partitioning with online locality-band routing and a KV-block–coindexed signature cache.
Compiles natural-language function specifications into compact, locally-executable neural programs (PAW) that run on a small frozen interpreter; a 4B compiler emits LoRA adapters for a 0.6B runtime to provide offline, low-memory fuzzy text functions.
Provides a large Mixture-of-Experts instruct LLM (295B total parameters, 21B active, 256K context) optimized for reasoning, long-context retention and agent workflows; open-sourced under Apache-2.0.
Teaches how to build, evaluate, and train LLM agents through three implementation assignments and a research project. CMU's Fall 2026 course connects tool use, memory, and planning with SFT, reinforcement learning, and agent safety.
Generates temporally grounded captions for dense multi-event videos by restructuring autoregressive token dependencies to enable lossless parallel decoding; introduces a latent global planning module and event-factorized parallel decoding to improve grounding accuracy and achieve large decoding speedups.
Provides GGUF-quantized local-deploy weights for a 1B MiniCPM5-derived conversational LLM, embedding a 'thinking' chat template and supporting up to 128K-token context; ships Q4/Q5/Q8/F16 quant files (Q8_0 recommended) for llama.cpp, Ollama, and LM Studio.
A 1B-parameter 'Thinking' language model fine-tuned on Fable 5 to improve coding and instruction-following; supports chain-of-thought style outputs, XML tool-call format, and up to 128K-token context, with GGUF builds for single-GPU local deployment.
Proposes SkillOpt-Lite, a minimal pipeline for optimizing LLM agent skills by treating rollout traces as filesystem files and applying trajectory exploration, consensus mining, and independent validation; integrates as a one-line VSCode Copilot command and reports cross-benchmark improvements that let smaller models sometimes outperform larger ones.
Instruction-tuned compact conversational model (Qwen3-4B-based) that generates short, chat-style replies and is optimized to run on a single mid-range GPU. Uses ChatML prompts, bfloat16 safetensors and is released under Apache-2.0; the model card notes a joke/placeholder disclaimer.