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 an initial experiment done on the at-this-point-infamous Llama 3.1 8B model, in an attempt to retain its smartness while addressing its abysmal lack of imagination/creativity. As always, feedback is welcome, and begone if you demand perfection.
The second stage, like the Sunfall series, follows the Silly Tavern preset, so ymmv in particular if you use some other tool and/or preset.
Feedback so far: The model is a bit less careful (aka sloppy) with things like formatting, potentially due to the overly impressionable pretraining part. I am adjusting the approach a bit and will post an update.
Parameter suggestions:
I did all my testing with temp 1, min-p 0.1, DRY 0.8. I enabled XTC at higher contexts.
Training details:
- Stage 1 (continued pretraining)
- Target: meta-llama/Llama-3.1-8B (resulting LoRA merged into meta-llama/Llama-3.1-8B-Instruct)
- LoRA dropout 0.5 (motivation)
- LoRA rank 64, alpha 128 (motivation)
- LR constant 1e-5
- LoRA+ with LR Ratio: 5
- Context size: 16384
- 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 1.5e-5 (min 1.5e-6)
- LoRA+ with LR Ratio: 5
- Context size: 16384
- Gradient accumulation steps: 4
- Epochs: 1
Initially I tried to keep stage 2 ultra thin/light to only minimally repair the merge-damage and trying to retain the base model's IQ, but this did not work out well, so I ultimately ramped up the dataset to include external sources (see dataset metadata) to boost the RP capacity of the model.
- Downloads last month
- 20