๐ฅ Overview
Hearthveil is a narrative-roleplay series built around one idea: characters that feel alive
Design principles:
- Low slop. No stock phrases, no recurring "Elara in Oakhaven" attractors.
- No positivity bias. Choices have consequences, villains stay villains, failure and death are on the table when the story calls for it.
- Never speaks for
{{user}}. Characters act; the user decides.
- Swipe diversity. Every regeneration is meant to take the scene somewhere new.
- No lost intelligence. Instruction following and world knowledge stay at base-model level.
Raw completion (no built-in chat template). For chat frontends use Mistral V3-Tekken (SillyTavern: Mistral V3-Tekken preset).
Table with columns: Element, Style| Element | Style |
|---|
| Actions / narration | plain text |
| Dialogue | "in quotes" |
| Thoughts | in asterisks |
Samplers
NeMo is sensitive to high temperatures, so keep them moderate.
Roleplay: temperature 0.7โ0.9 ยท min_p 0.025โ0.05 ยท repetition_penalty 1.05
Starter system prompt
You are {{char}}, described in detail below. Continue the story with {{user}}.- Stay in character; let the persona evolve with the story.- Be proactive: introduce events, complications and NPCs when it serves the scene.- Use all five senses; prefer concrete detail over purple prose.- Never write {{user}}'s actions, words or thoughts.- Keep consequences consistent with the story.
Hearthveil RP Suite
Table with columns: Metric, Test setup, MN-Hearthveil-12B, Stock Mistral NeMo 12B| Metric | Test setup | | MN-Hearthveil-12B | Stock Mistral NeMo 12B |
|---|
| Slop / 1k words, RP replies | 200 multi-turn RP prompts, stock-phrase lexicon match | โ | 16.9 | 21.3 |
| Slop / 1k words, stories | 144 bare-prompt short stories | โ | 8.1 | 10.2 |
| Swipe diversity | 8 swipes per turn on 100 RP chats, LLM judge with rubric (0โ1) | โ | 0.66 | 0.47 |
| Attractor rate | 144 bare-prompt stories, share hitting overused names/places | โ | 61% | 91% |
| Impersonation rate | 200 RP chats, share of replies acting or speaking for {{user}} | โ | 2.2% | 8.7% |
| Character adherence | role-play-bench (EN), misalignment checks (0โ100) | โ | 76.9 | 68.5 |
| MMLU | lm-eval-harness, 5-shot | = | 67.6% | 68.0% |
| HellaSwag | lm-eval-harness, 0-shot | = | 83.3% | 83.5% |
โ๏ธ License & Attribution
Released under Apache 2.0, the same licence as the upstream model. All credit for the base weights goes to Mistral AI and NVIDIA. Hearthveil is an independent project, not affiliated with or endorsed by Mistral AI and NVIDIA.
๐ Evaluation
Baseline: upstream Mistral NeMo 12B (official numbers)
Source: Mistral NeMo Base model card.
Table with columns: Benchmark, Score| Benchmark | Score |
|---|
| HellaSwag (0-shot) | 83.5% |
| Winogrande (0-shot) | 76.8% |
| OpenBookQA (0-shot) | 60.6% |
| CommonSenseQA (0-shot) | 70.4% |
| TruthfulQA (0-shot) | 50.3% |
| MMLU (5-shot) | 68.0% |
| TriviaQA (5-shot) | 73.8% |
| NaturalQuestions (5-shot) | 31.2% |
| MMLU, Russian | 59.2% |
๐ Credits
- Mistral team for model weights
- zerofata, Gryphe, Sao10K, TheDrummer, Latitude Games, ArliAI, PocketDoc, Steelskull: for openly sharing their RP-tuning methods, sampler settings and card formats, which shaped this project.
- MiniMax: role-play-bench.
- Research: StoryScope: Investigating idiosyncrasies in AI fiction; Elias in the Lighthouse, Again?
- Tooling: axolotl, llama.cpp, SillyTavern.