AI Dataset2026
Provides 1.7M agent interaction traces in terminus-2 format for training and evaluating agentic LLMs and RL agents. Compiled from 219 source datasets across code repair, shell, math, competitive programming and general tasks; produced with the Harbor harness.
AI Dataset2026
Aggregates 750k+ Harbor-compatible agentic tasks from 100+ public sources (Parquet shards preserved). Includes tasks with and without verifiers for RL evaluation or SFT/datagen workflows, enabling reproducible trace generation.
Preview of an MoE model family (V4-Pro: 1.6T params, 49B active; V4-Flash: 284B, 13B active) built for 1M-token contexts. A hybrid attention design cuts single-token inference FLOPs to 27% and KV cache to 10% versus V3.2 at million-token length.
AI Dataset2026
Provides tick-aligned Counter-Strike 2 player POV video clips with per-tick inputs and world-state sidecars — near-lossless 1280×720@32fps video, per-player stereo audio, and parquet indexes for event/kill/round filtering; suited for RL, video classification and clip mining.
AI Dataset2026
Multimodal STEM problem set for verifiable, answer-supervised training and RL: contains single-image, multi-panel, and multi-image PhD-level questions across physics, math, chemistry and biology. Each example has a deterministic ground-truth answer, enabling reward modeling and automated evaluation.
AI Dataset2026
Pairs natural-language instructions with executable setup artifacts and Python reward functions to create verifiable computer-use agent tasks. Provides a Parquet task table for fast filtering plus a compressed archive of runnable task bundles; several web task endpoints are placeholders that require a local CUA-Gym-Hub deployment.
AI Model2026
Research-focused text-to-image foundation model that prioritizes training efficiency: a 3.8B-parameter architecture trained on an 800M image-text corpus with mixed-resolution learning, FLUX.2 VAE, RL tuning, and a distilled 4-step Lens-Turbo for fast high-resolution generation.
AI Dataset2026
RL training dataset for long-context language-model fine-tuning with ~23K samples and nine reward types, provided in Parquet with bilingual ground-truth and reward metadata for direct RL/bench evaluation.
Combines internalizing general skills with task-specific skill utilization via a difficulty-aware router to improve in-distribution and out-of-distribution performance for agentic RL. Uses privileged distillation for hard tasks and diagnostic probing for easy tasks; evaluated on ALFWorld and WebShop.
Uses search-agent reading traces and tiered distractors to train LLMs for long-context, multi-hop reasoning, and introduces a rubric reward that supervises entity-level steps (applied only to correct finals). Improves evidence-grounded reasoning and resists reward hacking across 4B–30B models.
Proposes TrOPD, a method that restricts token-level on-policy distillation to regions where teacher supervision is reliable to stabilize training under teacher–student distribution mismatch. Adds outlier handling (clipping, masking, forward-KL) and off-policy guidance; shows consistent gains on math reasoning, code generation and general benchmarks.
Analyzes how single-domain RL fine-tuning on LLMs induces cross-domain interference and shows this damage concentrates in a low-dimensional shared conflict subspace; proposes a local perturbation theory and short domain "refresh" procedures that selectively recover earlier domains with minimal collateral loss.