hflog/grimjim-fireblossom-32K-7B

Fireblossom-32K-7B

This is a merge of pre-trained language models created using mergekit.

For this merge, I went back to Mistral 7B v0.1 for the literal base model for task arithmetic merger, which can be pushed to at least 16K context length after adjusting rope theta from 10K to 100K. With the original (true) base model, the models merged in should be mathematically equivalent to LoRA adapters. I left the original 32K context claimed by Mistral 7B v0.1.

The goal was a merge model more varied in its outputs, a goal which inherently harms accuracy in favor of creativity. To this end, I chose a model trained to be strong at narrative roleplay (cgato's work) along with three models that were good at reasoning (fine-tunes by HuggingFaceH4 and SanjiWatsuki). The result appears to be good at following card instructions, perhaps to a fault.

Sampler settings: Tested lightly with temperature=0.7 and minP=0.01. For greater creativity, boost temperature.

Prompts: Alpaca format natively supported, although ChatML was used during testing.

Download options:

• full weights
• Q8_0 GGUF
• 8.0bpw h8 exl2

Merge Details

Merge Method

This model was merged using the task arithmetic merge method using mistralai/Mistral-7B-v0.1 as a base.

Models Merged

The following models were included in the merge:

• HuggingFaceH4/zephyr-7b-beta
• cgato/TheSpice-7b-v0.1.1
• SanjiWatsuki/Kunoichi-DPO-v2-7B
• SanjiWatsuki/Kunoichi-7B

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: mistralai/Mistral-7B-v0.1
    # no parameters necessary for base model
  - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
    parameters:
      weight: 0.45
  - model: cgato/TheSpice-7b-v0.1.1
    parameters:
      weight: 0.05
  - model: HuggingFaceH4/zephyr-7b-beta
    parameters:
      weight: 0.05
  - model: SanjiWatsuki/Kunoichi-7B
    parameters:
      weight: 0.45
merge_method: task_arithmetic
base_model: mistralai/Mistral-7B-v0.1
dtype: float16