---
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
---

![slush.jpg](https://huggingface.co/crestf411/L3.1-8B-Slush-v1.1/resolve/main/slush.jpg?)

([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
```