A fine‑tuned llm‑jp‑4‑8b‑instruct model that removes censorship using Arbitrary‑Rank Ablation with specified parameters, achieving a KL divergence of 0.0781 and only 5/100 refusals, built on the LLM‑jp‑4 series trained via SFT without DPO or RL.
Description in other languages
Statistics
Statistics as of Nov 16, 2026
Hugging Face downloads (last 30 days)
1,000
Hugging Face likes
13
Listed
Nov 10, 2026
Popular in the same category
- japanese-gpt-neox-smallrinna656kA 12‑layer, 768‑hidden Japanese GPT‑NeoX model trained on CC‑100, C4, and Wikipedia, compatible with Huggingface, with an optional toy prefix‑tuning weight that forces each sentence to end with a smiling‑face emoji.
- Spark-X2.5-4B-Heretic-jp-ggufsoyaakinohara589kJapanese‑adapted Spark‑X2.5‑4B model (Heretic JP) with refusal attenuation, pretraining, instruction tuning and tool‑calling repair, available in GGUF quantizations from BF16 to Q4_K_M (~2.5 GiB), usable via llama.cpp for 1,048,576‑token context.
- llm-jp-4-33b-thinking-ggufllm-jp271kLLM‑jp‑4‑33b‑thinking‑gguf is a 33 B Japanese language model from the National Institute of Informatics, pre‑trained with SFT and DPO (no RL), available for llama.cpp via an LLM‑jp fork, and includes detailed usage instructions in its cookbook.
- llm-jp-4.1-8b-thinkingllm-jp80kLLM‑jp‑4.1‑8b‑thinking is a 33 B transformer language model from NII’s LLM R&D Center, pre‑trained with mid‑training only and aligned via SFT/DPO (no RL), designed for Japanese NLP tasks and documented in a cookbook and technical blog.
- llm-jp-3-150mllm-jp75kLLM‑jp‑3‑150m—a 150‑M‑parameter Japanese language model from the National Institute of Informatics’ LLM R&D Center—is distributed in Hugging Face Transformers format, requiring torch ≥ 2.3.0, transformers ≥ 4.40.1, accelerate ≥ 0.29.3, flash‑attn ≥ 2.5.8, and was pretrained on Japanese Wikipedia, Common Crawl, WARP/PDF, WARP/HTML, and Kaken data using a unigram byte‑fallback tokenizer.
- llm-jp-4-8b-instructllm-jp49kllm‑jp‑4‑8b‑instruct is a 4.1 B‑parameter Japanese LLM from NII’s LLM‑jp‑4 series, pretrained on large corpora and then fine‑tuned solely with supervised instruction data (no DPO/REINFORCE) and accompanied by cookbook‑style usage guides and a byte‑fallback unigram tokenizer.
Resources in similar research fields
chatvrmpixiv848ChatVRM is a demo application that allows you to easily chat with 3D characters in your browser.GitHubShared: Large language models- sarashina2.2-3b-instruct-v0.1sbintuitions16kOffers an autoregressive Japanese language model (sarashina2.2‑3B‑instruct‑v0.1) from SB Intuitions, benchmarked against other models, with example usage scripts and a note that safety training is limited.Shared: Large language models
aiavatarkituezo704Building AI-based conversational avatars lightning fastGitHubShared: Large language models
OneCompressionFujitsuResearch445Post-training quantization (PTQ) pipeline for LLM developed by Fujitsu Laboratories. Integrated QEP (NeurIPS 2025), ILP mixed precision, pre-rotation processing, and vLLM plugin. Paper: arXiv:2603.28845.GitHubShared: Large language models- llm-jp-4-8b-basellm-jp11kA repository that hosts the 8.6 B‑parameter llm‑jp‑4‑8b‑base transformer from the National Institute of Informatics’ LLM R&D Center, trained with pre‑training and mid‑training followed by supervised fine‑tuning and direct preference optimization (no reinforcement learning), and provides a PyTorch‑transformers usage guide.Shared: Large language models
- Llama-3-Swallow-8B-Instruct-v0.1tokyotech-llm9kLlama3 Swallow is a Japanese‑enhanced Meta Llama 3 family released July 1 2024, offering 8B and 70B variants in Instruct and chat forms fine‑tuned with SFT and Chat Vector on Megatron‑LM, and benchmarked on key Japanese NLP tasks.Shared: Large language models