File size: 3,311 Bytes
59e15f2 20cf5d4 59e15f2 777bc6a 59e15f2 d791298 59e15f2 a0ebd13 d791298 59e15f2 |
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 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 |
---
library_name: transformers
tags:
- not-for-all-audiences
- mergekit
datasets:
- crestf411/LimaRP-DS
- Gryphe/Sonnet3.5-Charcard-Roleplay
- anthracite-org/c2_logs_32k_mistral-v3_v1.2_no_system
- anthracite-org/kalo-opus-instruct-22k-no-refusal-no-system
- anthracite-org/kalo-opus-instruct-3k-filtered-no-system
- anthracite-org/nopm_claude_writing_fixed
base_model:
- Qwen/Qwen2.5-32B-Instruct
---
([GGUFs](https://huggingface.co/mradermacher/Q2.5-32B-Slush-i1-GGUF))
**Slush** is a two-stage model trained with high LoRA dropout, where stage 1 is a pretraining continuation on the base model, aimed at boosting the model's creativity and writing capabilities. This is then merged into the instruction tune model, and stage 2 is a fine tuning step on top of this to further enhance its roleplaying capabilities and/or to repair any damage caused in the stage 1 merge.
This is still early stage. As always, feedback is welcome, and begone if you demand perfection.
The second stage, like the *Sunfall* series, follows the Silly Tavern preset (ChatML), so ymmv in particular if you use some other tool and/or preset.
## Parameter suggestions
I did all my testing with temp 1, min-p 0.1, DRY 0.8, but enabled XTC as context grew and/or the model started saying "the same stuff".
Qwen 2.5 32B Instruct (vanilla) has a strong tendency to start speaking for the user, especially in narrator scenarios. I was unable to properly train this out of the model completely, so you may want to add e.g. "\nYou" as a stopping string, and enable "trim incomplete sentences", especially if you have banned sentences.
The model has a tendency to add an unnecesary final paragraph to responses during roleplay, sort of like a "summary" of how the character is feeling. Keeping it is OK, but it may snowball quickly. Hoping to address this but not sure how.
## Training details
* Stage 1 (continued pretraining)
* Target: Qwen/Qwen2.5-32B (resulting LoRA merged into Qwen/Qwen2.5-32B-Instruct)
* LoRA dropout 0.5 ([motivation](https://arxiv.org/abs/2403.00946))
* LoRA rank 32, alpha 64 ([motivation](https://arxiv.org/abs/2410.21228))
* LR cosine 4e-6
* [LoRA+](https://arxiv.org/abs/2402.12354) with LR Ratio: 15
* Context size: 8192
* Gradient accumulation steps: 4
* Epochs: 1
* Stage 2 (fine tune)
* Target: Stage 1 model
* LoRA dropout 0.5
* LoRA rank 32, alpha 64
* LR cosine 5e-6 (min 5e-7)
* [LoRA+](https://arxiv.org/abs/2402.12354) with LR Ratio: 15
* Context size: 16384
* Gradient accumulation steps: 4
* Epochs: 1
## Merge Details
### Merge Method
This model was merged using the [TIES](https://arxiv.org/abs/2306.01708) merge method.
### Configuration
The following YAML configuration was used to produce this model:
```yaml
models:
- model: stage1-model
parameters:
weight: 1
density: 1
- model: stage2-model
parameters:
weight: 1
density: 1
- model: Qwen/Qwen2.5-32B-Instruct
parameters:
weight: 0.9
density: 0.9
merge_method: ties
base_model: Qwen/Qwen2.5-32B
parameters:
weight: 0.9
density: 0.9
normalize: true
int8_mask: true
tokenizer_source: Qwen/Qwen2.5-32B-Instruct
dtype: bfloat16
```
|