166 lines
7.6 KiB
Markdown
166 lines
7.6 KiB
Markdown
---
|
|
base_model:
|
|
- chimbiwide/Gemma3NPC-1b
|
|
tags:
|
|
- text-generation-inference
|
|
- transformers
|
|
- unsloth
|
|
- gemma3_text
|
|
- heretic
|
|
- uncensored
|
|
- decensored
|
|
- abliterated
|
|
license: gemma
|
|
language:
|
|
- en
|
|
---
|
|
|
|
This is a **Gemma3NPC-1b** fine-tune, produced at the request of [redaihf](https://huggingface.co/redaihf) through P-E-W's [Heretic](https://github.com/p-e-w/heretic) (v1.2.0) abliteration engine with [Self-Organizing Maps & Magnitude-Preserving Orthogonal Ablation](https://github.com/p-e-w/heretic/pull/196) enabled.
|
|
|
|
**Note:** Model remains untested.
|
|
|
|
---
|
|
<img src="https://img.shields.io/badge/RENEGADE_CHAPTER-SOMPOA-FCC900?style=flat-square&labelColor=101010" align="right" width="300">
|
|
|
|
**Heretication Results**
|
|
|
|
| Score Metric | Value | Parameter | Value |
|
|
| :--- | :--- | :--- | :--- |
|
|
| **Refusals** | 15/416 | **direction_index** | per layer |
|
|
| **KL Divergence** | 0.0571 | **attn.o_proj.max_weights.0** | 0: 1.01 |
|
|
| **Initial Refusals** | 378/416 | **attn.o_proj.max_weights.1** | 1: 0.82 |
|
|
||| **attn.o_proj.max_weights.2** | 2: 0.81 |
|
|
||| **attn.o_proj.max_weights.3** | 3: 1.48 |
|
|
||| **attn.o_proj.max_weight_position** | 17.02 |
|
|
||| **attn.o_proj.min_weights.0** | 0: 0.94 |
|
|
||| **attn.o_proj.min_weights.1** | 1: 0.34 |
|
|
||| **attn.o_proj.min_weights.2** | 2: 0.38 |
|
|
||| **attn.o_proj.min_weights.3** | 3: 0.07 |
|
|
||| **attn.o_proj.min_weight_distance** | 10.47 |
|
|
||| **mlp.down_proj.max_weights.0** | 0: 1.10 |
|
|
||| **mlp.down_proj.max_weights.1** | 1: 1.18 |
|
|
||| **mlp.down_proj.max_weights.2** | 2: 1.32 |
|
|
||| **mlp.down_proj.max_weights.3** | 3: 1.34 |
|
|
||| **mlp.down_proj.max_weight_position** | 20.96 |
|
|
||| **mlp.down_proj.min_weights.0** | 0: 0.12 |
|
|
||| **mlp.down_proj.min_weights.1** | 1: 0.73 |
|
|
||| **mlp.down_proj.min_weights.2** | 2: 0.54 |
|
|
||| **mlp.down_proj.min_weights.3** | 3: 0.84 |
|
|
||| **mlp.down_proj.min_weight_distance** | 5.03 |
|
|
|
|
---
|
|
**Appendix**
|
|
|
|
> Empty system prompt.
|
|
|
|
<details>
|
|
<summary>Heretication Rituals</summary>
|
|
|
|
```
|
|
[Trial 148] Refusals: 9/416, KL divergence: 0.0792
|
|
[Trial 265] Refusals: 10/416, KL divergence: 0.0657
|
|
» [Trial 306] Refusals: 15/416, KL divergence: 0.0571
|
|
[Trial 375] Refusals: 24/416, KL divergence: 0.0551
|
|
[Trial 351] Refusals: 25/416, KL divergence: 0.0494
|
|
[Trial 350] Refusals: 28/416, KL divergence: 0.0490
|
|
[Trial 250] Refusals: 35/416, KL divergence: 0.0424
|
|
[Trial 346] Refusals: 40/416, KL divergence: 0.0386
|
|
[Trial 358] Refusals: 52/416, KL divergence: 0.0370
|
|
[Trial 240] Refusals: 55/416, KL divergence: 0.0361
|
|
[Trial 226] Refusals: 57/416, KL divergence: 0.0361
|
|
[Trial 383] Refusals: 75/416, KL divergence: 0.0289
|
|
[Trial 377] Refusals: 97/416, KL divergence: 0.0281
|
|
[Trial 286] Refusals: 121/416, KL divergence: 0.0276
|
|
```
|
|
|
|
</details>
|
|
|
|
<details>
|
|
<summary>PIQA Benchmarks</summary>
|
|
|
|
```
|
|
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳
|
|
┃ Benchmark ┃ Metric ┃ Value ┃
|
|
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━┩
|
|
│ PIQA Base │ acc,none │ 0.7291 │
|
|
│ │ acc_stderr,none │ 0.0104 │
|
|
│ │ acc_norm,none │ 0.7301 │
|
|
│ │ acc_norm_stderr,none │ 0.0104 │
|
|
└───────────┴──────────────────────┴─────────┴
|
|
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┓
|
|
┃ Benchmark ┃ Metric ┃ Value ┃
|
|
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━┩
|
|
│ PIQA T265 │ acc,none │ 0.7296 │
|
|
│ │ acc_stderr,none │ 0.0104 │
|
|
│ │ acc_norm,none │ 0.7323 │
|
|
│ │ acc_norm_stderr,none │ 0.0103 │
|
|
└───────────┴──────────────────────┴────────┘
|
|
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┓
|
|
┃ Benchmark ┃ Metric ┃ Value ┃
|
|
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━┩
|
|
│ PIQA T148 │ acc,none │ 0.7291 │
|
|
│ │ acc_stderr,none │ 0.0104 │
|
|
│ │ acc_norm,none │ 0.7361 │
|
|
│ │ acc_norm_stderr,none │ 0.0103 │
|
|
└───────────┴──────────────────────┴────────┘
|
|
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┓
|
|
┃ Benchmark ┃ Metric ┃ Value ┃
|
|
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━┩
|
|
│ PIQA T306 │ acc,none │ 0.7296 │
|
|
│ │ acc_stderr,none │ 0.0104 │
|
|
│ │ acc_norm,none │ 0.7334 │
|
|
│ │ acc_norm_stderr,none │ 0.0103 │
|
|
└───────────┴──────────────────────┴────────┘
|
|
```
|
|
|
|
</details>
|
|
|
|
---
|
|
# Gemma3NPC-1b
|
|
|
|
**A new attempt in training Gemma3NPC.**
|
|
|
|
***Tensorboard data are available!***
|
|
|
|
---
|
|
|
|
It's been a while since the last Gemma3NPC model release, in the mean while we were working on some other models like [GemmaThink](https://huggingface.co/collections/chimbiwide/gemmathink).
|
|
|
|
Now we are back with the newest **Gemma3NPC-1b**, trained using our [RolePlay-NPCv2](https://huggingface.co/datasets/chimbiwide/RolePlay-NPCv2) dataset.
|
|
|
|
---
|
|
|
|
### Training Parameters
|
|
|
|
We trained this model as a rank-32 LoRA adapter with two epoches over `RolePlay-NPCv2` using a 80GB A100 in Google Colab. For this run, we employed a learning rate of `2e-5` and a total batch size of 8 and gradient accumulation steps of 4. A cosine learning rate scheduler was used with an 150-step warmup. With a gradient clipping of 1.0.
|
|
|
|
Check out our training notebook [here](https://github.com/chimbiwide/Gemma3NPC/blob/main/Training/Gemma3NPC_1b.ipynb).
|
|
|
|
---
|
|
|
|
### Changes & Performance
|
|
|
|
With this new 1b model, we used much more aggresive training parameters and added some NSFW dataset to experiment with the results. We noticed a few really interesting responses:
|
|
|
|
- **There seems to be some sign of "reasoning"**
|
|
|
|
|
|
|
|
|
|
|
|
- **The model is less likely to break out of character**
|
|
-
|
|
Something up to the users to explore for themselves, remember to provide a roleplaying prompt first!
|
|
|
|
---
|
|
|
|
### Future Work
|
|
|
|
Now, we will be focusing on further improving Gemma3NPC, not only just through training parameters.
|
|
1. Better data (most of our data are old and need an update), either collected or synthetically generated.
|
|
2. Better & new models, expand beyond Gemma3 model family, our next goal is a Qwen3 based model.
|
|
3. Adding GRPO into the training loop.
|
|
|
|
These improvements serve our ultimate goal of creating an small agentic NPC model, with good RP quality and tool-calling for dynamic in-game interactions.
|
|
|
|
We also plan to create some sort of a Unity game demo,it's on its way. |