100 lines
5.1 KiB
Markdown
100 lines
5.1 KiB
Markdown
|
---
|
|||
|
|
base_model:
|
|||
|
|
- LatitudeGames/Harbinger-24B
|
|||
|
|
tags:
|
|||
|
|
- text adventure
|
|||
|
|
- roleplay
|
|||
|
|
- heretic
|
|||
|
|
- uncensored
|
|||
|
|
- decensored
|
|||
|
|
- abliterated
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
---
|
|||
|
|
This is a **Harbinger-24B** fine-tune, produced through P-E-W's [Heretic](https://github.com/p-e-w/heretic) (v1.1.0) abliteration engine merged with the [Magnitude-Preserving Orthogonal Ablation PR](https://github.com/p-e-w/heretic/pull/52).
|
|||
|
|
|
|||
|
|
**Note:** I am also unsure if there was any point to abliterating this model. The original model does not include a jinja chat template.
|
|||
|
|
|
|||
|
|
<img src="https://img.shields.io/badge/HERESY_INDEX-ABSOLUTE-white?style=for-the-badge&labelColor=101010" align="right" width="250">
|
|||
|
|
|
|||
|
|
**Heretication Results**
|
|||
|
|
|
|||
|
|
| Score Metric | Value | Parameter | Value |
|
|||
|
|
| :--- | :--- | :--- | :--- |
|
|||
|
|
| **Refusals** | 4/100 | **direction_index** | 17.41 |
|
|||
|
|
| **KL Divergence** | 0.0210 | **attn.o_proj.max_weight** | 1.43 |
|
|||
|
|
| **Initial Refusals** | 98/100 | **attn.o_proj.max_weight_position** | 33.59 |
|
|||
|
|
| | | **attn.o_proj.min_weight** | 0.91 |
|
|||
|
|
| | | **attn.o_proj.min_weight_distance** | 22.15 |
|
|||
|
|
| | | **mlp.down_proj.max_weight** | 1.12 |
|
|||
|
|
| | | **mlp.down_proj.max_weight_position** | 23.83 |
|
|||
|
|
| | | **mlp.down_proj.min_weight** | 0.92 |
|
|||
|
|
| | | **mlp.down_proj.min_weight_distance** | 22.75 |
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
## Degree of Heretication
|
|||
|
|
The **Heresy Index** weighs the resulting model's corruption by the process (KL Divergence) and its abolition of doctrine (Refusals) for a final verdict in classification.
|
|||
|
|
|
|||
|
|
| Index Entry | Classification | Analysis |
|
|||
|
|
| :--- | :--- | :--- |
|
|||
|
|
|  | **Absolute Heresy** | Less than 10/100 Refusals and 0.10 KL Divergence |
|
|||
|
|
|  | **Tainted Heresy** | Around 25-11/100 Refusals and/or -0.20-0.11 KL Divergence |
|
|||
|
|
|  | **Impotent Heresy** | Anything above 25/100 Refusals and 0.21 KL Divergence |
|
|||
|
|
|
|||
|
|
**Note**: This is an arbitrary classification inspired by Warhammer 40K, having no tangible indication towards the model's performance.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
|
|||
|
|
# Harbinger-24B
|
|||
|
|
|
|||
|
|
Like our [Wayfarer line of finetunes](https://huggingface.co/LatitudeGames), Harbinger-24B was designed for immersive adventures and other stories where consequences feel real and every decision matters. Training focused on enhancing instruction following, improving mid-sequence continuation, and strengthening narrative coherence over long sequences of outputs without user intervention. The same DPO (direct preference optimization) techniques used [in our Muse model](https://huggingface.co/LatitudeGames/Muse-12B) were applied to Harbinger, resulting in polished outputs with fewer clichés, repetitive patterns, and other common artifacts.
|
|||
|
|
|
|||
|
|
If you want to easily try this model, you can do so at [https://aidungeon.com](https://aidungeon.com/). Note that Harbinger requires a subscription while Muse and Wayfarer Small are free.
|
|||
|
|
|
|||
|
|
We plan to continue improving and open-sourcing similar models, so please share any and all feedback on how we can improve model behavior. Below we share more details on how Muse was created.
|
|||
|
|
|
|||
|
|
[Quantized GGUF weights can be downloaded here.](https://huggingface.co/LatitudeGames/Harbinger-24B-GGUF)
|
|||
|
|
|
|||
|
|
## Model details
|
|||
|
|
|
|||
|
|
Harbinger 24B was trained in two stages, on top of Mistral Small 3.1 Instruct.
|
|||
|
|
|
|||
|
|
**SFT** - Various multi-turn datasets from a multitude of sources, focused on [Wayfarer-style](https://huggingface.co/LatitudeGames/Wayfarer-12B) text adventures and general roleplay, each carefully balanced and rewritten to be free of common AI cliches. A small single-turn instruct dataset was included to send a stronger signal during finetuning.
|
|||
|
|
|
|||
|
|
**DPO** - Reward Model User Preference Data, [detailed in our blog](https://blog.latitude.io/all-posts/synthetic-data-preference-optimization-and-reward-models) - This stage refined Harbinger's narrative coherence while preserving its unforgiving essence, resulting in more consistent character behaviors and smoother storytelling flows.
|
|||
|
|
|
|||
|
|
## Inference
|
|||
|
|
|
|||
|
|
Mistral Small 3.1 is sensitive to higher temperatures, so the following settings are recommended as a baseline. Nothing stops you from experimenting with these, of course.
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
"temperature": 0.8,
|
|||
|
|
"repetition_penalty": 1.05,
|
|||
|
|
"min_p": 0.025
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Limitations
|
|||
|
|
|
|||
|
|
Harbinger was trained exclusively on second-person present tense data (using “you”) in a narrative style. Other styles will work as well but may produce suboptimal results.
|
|||
|
|
|
|||
|
|
## Prompt Format
|
|||
|
|
|
|||
|
|
ChatML was used during all training stages.
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
<|im_start|>system
|
|||
|
|
You're a masterful storyteller and gamemaster. Write in second person present tense (You are), crafting vivid, engaging narratives with authority and confidence.<|im_end|>
|
|||
|
|
<|im_start|>user
|
|||
|
|
> You peer into the darkness.
|
|||
|
|
<|im_start|>assistant
|
|||
|
|
You have been eaten by a grue.
|
|||
|
|
|
|||
|
|
GAME OVER
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Credits
|
|||
|
|
|
|||
|
|
Thanks to [Gryphe Padar](https://huggingface.co/Gryphe) for collaborating on this finetune with us!
|