A distilled LoRA adapter for Qwen-Image-2.1 that runs text-to-image generation and instruction-driven image editing in a few transformer passes (shipped as a 6-step r256 LoRA). Samples with a fixed sigma schedule, no classifier-free guidance; non-commercial research license.
A mobile-planning agent paper that develops a closed-loop AI-for-AI lifecycle to generate data, train a planner model, and co-evolve a runtime Harness for multi-app task execution. Demonstrates top performance on MobilePA-Bench with improved tool use, memory, skill coordination and low estimated per-task output cost.
Shows that private post-training changes leave measurable "behavioral shadows" in task-unrelated single-word outputs and proposes Active Taskless Distillation (ATD) to transfer capabilities to a student using only those single-word teacher responses, without teacher logits or parameters.
Provides compact mixed-precision GGUF quantizations of UkisAI's Swift 1.5 (derived from Qwen3.8-27B), using ISTA GSQ-RCO per-tensor allocations with Swift-specific refinement. Offers multiple 8–12 GB tiers, optional MTP heads, and KLD evaluation against the Swift BF16 baseline.
Compresses Qwen3.8-27B into a 12.3 GB sensitivity-aware mixed-precision quantized checkpoint for long-horizon agent workloads; preserves BF16 fidelity (+0.02% PPL, 93.2% token Top‑1 agreement), supports 262K context and vLLM serving, text-only and Apache‑2.0 licensed.
Capability-targeted compression of Qwen3.8-Flash-Next: half the experts are removed and remaining weights quantized to 3.5 bpw, producing a 58.4 GB GGUF (29.6 GB resident) that preserves coding and multimodal ability while trading off other domains.
Provides a high-fidelity mixed-precision (≈3-bit) GGUF quant of Qwen3.8-27B tailored for long-horizon agent workloads and cyber-focused red-teaming. Preserves reasoning, thinking mode, MTP speculative decoding and vision (via a separate mmproj); released as an uncensored/abliterated research build under Apache-2.0.
Investigates how rollout policy, token-level KL direction, and learning rate each affect LLM distillation across Llama3 and Qwen2.5 on reasoning tasks; finds KL direction and learning rate dominate outcomes while rollout policy has a modest effect.
Collection of 1.44M unique Turkish voice‑assistant sentences (≈1,819 hours estimated), normalized for TTS and organized by service domains (appointments, banking, e‑commerce). Designed for training and evaluating TTS and text-generation models; licensed CC BY 4.0 with required attribution.
Defines and evaluates AREX-2, an LLM agent that iteratively self-improves at test time via reflection and long-horizon execution. Trained on long-horizon improvement trajectories from ML engineering and algorithmic programming (built on Qwen3.8-27B), it scales with more rounds and achieves strong benchmark scores.
Demonstrates that pretrained transformers typically use only ~1–3 lines of depth to follow reference chains, and that a task‑trained rank‑8 LoRA applied at one early layer (with all other weights frozen) can extend reference‑following to dozens or hundreds of lines while adding only a few ten‑thousand parameters.
Compresses recurrent states in linear-attention LLMs via spatial–temporal post-training quantization, allocating bits by error lifetime and per-row impact to preserve accuracy (6-bit ≈ FP32) while cutting serving memory up to 68.7%.