How To · 3
Anti-H 2.0 is our LoRA for Mistral Nemo. It is published on Hugging Face; this page explains what it is and how to use it.
The LoRA
A small add-on that keeps a companion in its own lane (no writing your words, actions or thoughts) and keeps tense and quotation marks clean. Trained for Mistral Nemo. For each training category I hand-wrote about 400 scenarios, expanded them with open models, and hand-edited the results.
Get Anti-H 2.0 on Hugging FaceOn Hugging Face under WordBlockLabs, with a model card, the license and credits: huggingface.co/WordBlockLabs/solo-r-lane-lora. The repository is named solo-r-lane-lora; Anti-H 2.0 is the same adapter.
The repository holds the adapter in PEFT format, which vLLM and transformers load directly next to the base model. llama.cpp, KoboldCPP and Ollama need it converted to GGUF first; the model card has the starting points. A ready-made GGUF is not provided yet.
In four runs of 216 turns (432 turns per side, scripted player, three narration perspectives, three paragraph levels), head-hopping in what the player sees was 24% of replies for the base model and 6% with Anti-H 2.0, and tense stayed consistent in 99% of replies against 46%. It also found real costs: replies are shorter (119 words to 47), it repeats short phrases more (52 to 192 per 1,000 words), and it meets the paragraph minimum less often. A later freshness guard in the app cut the repeats to 72. It is a feasibility result, not a final claim.
Yes. If the model starts to head hop, it can be reset by stopping your LLM server, closing the SoloRoleplayer session and restarting both. The model is still being perfected and still hiccups now and again. It is not yet preventable and its rate is not yet measured.
Anti-H 1.x was the first LoRA, trained for a smaller Qwen model. Anti-H 2.0 replaces it.
A solo project. SoloRoleplayer// is for adults, 18 and over. It runs offline on your own computer and is not monitored: you are responsible for how you use it.