Tag
Explore by tags
Unifies video, audio, image and text understanding for enterprise Q&A, summarization, transcription and document intelligence. The NVFP4 quantized variant reduces footprint to ~20.9GB for more efficient single‑GPU deployment and is tuned for NVIDIA runtimes (vLLM, TensorRT).
Supervised fine-tuning dataset of 7,716 reasoning-focused Q&A examples distilled from the DeepSeek‑V4‑Flash teacher; provided as a cleaned JSONL train split for distillation and SFT experiments.
1,000 JSONL samples containing full chain-of-thought reasoning traces and final answers produced by DeepSeek‑V4‑Pro for use in student-model distillation and quality checks. Prompts sampled from Jackrong/GLM-5.1-Reasoning-1M-Cleaned; Apache‑2.0 licensed.
Collects real-world developer–AI coding sessions with full transcripts, tool calls, agent thinking traces, Git commits, and agent vs. human code attribution. Packaged as Parquet tables (conversations, sessions, commits, checkpoints, repositories) for analysis of agent behavior and human–AI collaboration.
Provides ~55K multimodal VQA items with matched contrastive pairs and model‑generated rationales across five categories (General, Reasoning, Math, Graph/Chart, OCR), enabling research on faithful visual reasoning and robustness. Train split: 54,844 examples; license unspecified—verify before use.
Provides unquantized BF16 weights of Qwen3.6-27B with the base model's MTP head grafted in for high-fidelity, uncensored text (and multimodal) generation. Includes deployment guidance and hardware-tuned variants for A100/H100 and Blackwell-class GPUs.
Contains full chain-of-thought traces and final answers generated by DeepSeek-V4-Pro for use as distillation supervision. Key features: full CoT exposure, ~1,000 mixed-domain samples (JSONL/Parquet), Apache-2.0 license — suitable for training student models but watch for source contamination.
Generates anime-style images from natural-language prompts with a full fine-tune family built on Z-Image Base — available as Base, 8-step and 4-step distillations, plus AIO and GGUF variants for 8GB/low-VRAM workflows (BF16/FP8 formats).
Provides multiple GGUF-quantized exports of Carnice V2 (a merged BF16 SFT of Qwen3.6-27B) optimized for llama.cpp and Hermes-style agent traces, with quant tiers targeted at 16–24GB local GPUs and agentic inference.
Cleaned dataset of reasoning-distillation examples derived from Claude Opus 4.7 outputs — 4,807 retained JSON chat rows after removing simulated-thinking, duplicates, and missing fields. Packaged for model distillation and reasoning evaluation; Apache-2.0 packaging with upstream Anthropic usage constraints.
GGUF-format, DS4-optimized quantized weights for DeepSeek-V4-Flash, offering q2 (≈80.8 GiB) and q4 (≈153.3 GiB) variants plus an optional small MTP file for speculative decoding. Built for the DS4 inference engine; MIT-licensed.
Training dataset for byte-level language identification across 334 languages with ~2.48M paragraph samples (primarily Wikipedia and open-licensed corpora). Curated to reduce multilingual contamination, boost low-resource coverage, target frequent confusions, and preserve per-row license metadata for attribution.