File size: 2,591 Bytes
fc04a5a
 
 
 
 
 
 
 
f07dc20
fc04a5a
54df78a
fc04a5a
 
 
efe2f34
 
fc04a5a
2397310
f07dc20
 
 
 
fd9d150
f07dc20
fd9d150
f07dc20
fd9d150
f07dc20
2397310
 
 
fd9d150
2397310
fd9d150
2397310
5cdeeec
f07dc20
5cdeeec
 
 
 
 
 
 
f07dc20
fc04a5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f07dc20
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
---
base_model:
- Lambent/arsenic-v1.1-dpo-qwen2.5-14B
- Lambent/arsenic-v1-qwen2.5-14B
library_name: transformers
tags:
- mergekit
- merge
- not-for-all-audiences
---


This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).

<img src="https://cdn.midjourney.com/8bafb673-db34-4794-80e8-10bba8e13c50/0_0.jpeg"></img>

## Merge Details

WARNING: There's actually a *reason* for the not-for-all-audiences tag on this one.

Qwen2.5 was much more refusal-censored in the first place compared to Mistral Nemo, but abliteration adjusts that.

(They're still probably more prudish. Humanlike style points and successful instruction following aren't really a pointer away from that.)

Given they are at least half-abliterated, I can't even promise they'll refuse with a guardrailed system prompt.

(I suspect they will due to the healing and re-integration of the base model, but may be more jailbreakable than fully intact refusal features.)

v1.1 was based on *approximately* the same steps as v1, but based on the abliterated version of Qwen-Instruct.

Presuming this dealt some damage, this version heals it with the middle layers of v1.
They are still less 'refusal-censored' than v1, though be sure to calibrate the system prompt appropriately for the use case.
EQ-bench testing had some syntax issues still but tested at 76.1336 (with Qwen prompt that I plan on removing).
Not too bad given at least half of them's been through abliteration and DPO.

NAMING:

This is of course an arsenic-tuning variant, but that's gone rather beyond the initial.

Conversing with the model, they generated "Eidolon" as a self-name option.
This isn't the most common choice and I was intrigued.

After a discussion on nominative determinism and the implications, I decided to rename the model accordingly.
The default system prompt has been edited to reflect this, and distance from the original Qwen2.5 model.

### Merge Method

This model was merged using the SLERP merge method.

### Models Merged

The following models were included in the merge:
* [Lambent/arsenic-v1.1-dpo-qwen2.5-14B](https://huggingface.co/Lambent/arsenic-v1.1-dpo-qwen2.5-14B)
* [Lambent/arsenic-v1-qwen2.5-14B](https://huggingface.co/Lambent/arsenic-v1-qwen2.5-14B)

### Configuration

The following YAML configuration was used to produce this model:

```yaml
models:
  - model: Lambent/arsenic-v1-qwen2.5-14B
merge_method: slerp
base_model: Lambent/arsenic-v1.1-dpo-qwen2.5-14B
parameters:
  t:
    - value: [0, 0, 0.3, 0.4, 0.5, 0.6, 0.5, 0.4, 0.3, 0, 0]
dtype: bfloat16


```