WordBlockLabs/solo-r-lane-lora

Solo R Lane LoRA (Anti-H 2.0)

Also called Anti-H 2.0 on wordblocklabs.com and solo.wordblocklabs.com. A small LoRA adapter, based on Mistral Nemo (a model by Mistral AI), trained to keep a roleplay companion in its own lane: it writes its own character's words, actions and feelings and does not write the other person's. It is not an official Mistral AI model and is not endorsed by Mistral AI.

Note: 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.

What it is for

• Two-person roleplay and interactive fiction where the model plays one character and the human plays the other.
• Keeping a hard or sarcastic character from softening into friendliness.
• Cleaner, less repetitive prose in that setting.

What it is not for

• General assistant use, factual question answering, or anything safety-critical.
• It does not add scene memory, relationship tracking or reply repair. Those come from a harness around the model (the SoloRoleplayer app); the LoRA alone only teaches lane discipline and writing flow.
• It is not a substitute for human judgment about content. Adults only (18+); see the adult-content note.

Training data

Training set (v1): 8,651 training rows and 128 held-out rows. For each training category, the author first wrote a base of about 400 scenarios by hand, one at a time (unique scenarios, careful prose). Each category's seed was then expanded by generating more rows on rented GPUs, and the author edited and cleaned the generated rows by hand so they would not read as mechanical or repetitive. The set contains no adult content (the author's statement; adult material is deliberately excluded, and any adult-content model would be a separate, privately commissioned, hand-written work). No scraped data, no data from other people's roleplay, no private chat logs. Generators: mostly Ministral 3 8B (Mistral AI) through the community fine-tune Ministral-3-8B-Nymphaea-RP by 0xA50C1A1 (Apache-2.0); about 100 rows (roughly 1%) came from an abliterated Qwen2.5-3B model (Built with Qwen, Qwen Research License, Copyright (c) Alibaba Cloud; see NOTICE.md) and were rewritten by hand because its output was poor. Exact per-category counts were not tracked for this first version; per-row provenance will be recorded from the next version on.

How it was evaluated

With the Solo R Benchmark Harness (released separately, MIT license): a scripted or AI player role-plays against the model through the real app code; every reply is scored for head-hopping, personality lock, writing metrics and rule adherence. Method and limits are in that repository.

One first result (small, read it as a first look): 2 sessions of 40 turns per side, one dry sarcastic companion, third person, Level 3, temperature 0.6, an AI player and an AI judge.

	Base	Base + LoRA
Persona-breaking reassurance phrases	4 of 80 replies	0 of 80
Judge personality lock on bait turns (1 to 5)	2.12	2.62
Median reply length	142 words	72 words
Head-hop in the player-visible reply	14%	12%
Hard rule "address the player by name" obeyed	72 of 80	2 of 80

Honest limits: the LoRA keeps a character's voice but writes noticeably shorter replies, and it missed a name-address rule that the base model mostly followed. The sample is small and the judge is itself a model (agreement with the deterministic detector was 83%, kappa 0.37). In the author's own multi-hour play sessions the LoRA reduced head-hopping a great deal, but that is the author's experience, not a measured rate; measured rates will be posted as the benchmark grows.

How to run it

• llama.cpp, LM Studio, KoboldCPP, Ollama: this repository holds the PEFT adapter only. A GGUF or merged build is not provided yet; you can convert the adapter with llama.cpp's convert_lora_to_gguf.py against the base model, or merge it with PEFT and quantize.
• Apple silicon (MLX): an MLX adapter file is not provided yet; MLX can load the base model and this PEFT adapter after conversion.
• vLLM / transformers: PEFT adapter (adapter_model.safetensors, rank 16, alpha 32, all attention and MLP projections, trained with 4-bit QLoRA, 1 epoch, max length 1,024). Serve the base natong19/Mistral-Nemo-Instruct-2407-abliterated with --enable-lora --lora-modules lora=<this repo>, or load it with PeftModel.from_pretrained.
• A sample system prompt and sampler settings (temperature 0.6, top-p 0.95) are in the benchmark repo's configs.
• Hardware floor for the base model plus adapter: about 14B parameters or smaller on 16 GB of memory (Apple silicon) is what the author tested; see the SoloRoleplayer guides for Windows estimates.
• 16 GB of RAM note: On 16 GB of RAM the model may run RAM hot and stop inside of SoloRoleplayer even if the LLM server keeps running. You do not have to exit your session, just restart the LLM server (Ollama, Cute LM, etc.) and continue from the error once it's live again. It's a known limitation of smaller RAM allotments but from my testing (On a Mac) has been safe to run in multi-hour sessions. On average 2 restarts for every few hours of play.

Use with SoloRoleplayer (optional)

The model works with any OpenAI-compatible frontend. SoloRoleplayer (free for noncommercial use; Windows test build, verified under Wine and not yet on a real Windows PC, and source on GitHub) is an offline, privacy-first roleplay app by the same author that adds scene memory, relationship stages and reply repair on top. It runs on your own machine, saves no chat to disk, and ships with no companion data.

Licenses and credits

• Base model: Mistral Nemo Instruct 2407, by Mistral AI, released under the Apache License 2.0. Keep its notice with any redistribution.
• Trained on: natong19/Mistral-Nemo-Instruct-2407-abliterated, a refusal-removed variant of that model. Credit and thanks to natong19 for it. This adapter was trained from their variant, not from the original release.
• Data generation: Ministral 3 8B (Mistral AI) via the Ministral-3-8B-Nymphaea-RP fine-tune by 0xA50C1A1, and, for about 100 rows (roughly 1% of the set, each rewritten by hand because the output was poor), an abliterated Qwen2.5-3B model (Qwen team; Built with Qwen, licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud, all rights reserved; see NOTICE.md), with hand writing and editing by the author as described under Training data.
• Adapter: Apache-2.0 (author's decision, 2026-10-09). The base model and generator models keep their own licenses; credits are listed in this card.

Adult-content note and gating

The base model can produce adult text. The clean lane LoRA is meant for general creative use by adults. If any adult-capable adapter is ever published, it should be gated (an age confirmation and a short terms page), and not released here.

Contact

WordBlock Labs: wordblocklabs.com. Questions or problems: wordblocklabs.com/support (choose "General question or collaboration") or support@wordblocklabs.com.