MN-Tiramisu-12B

MN-Tiramisu-12B

This is a really yappity-yappy yapping model that's good for long-form RP. Tried to rein it in with Mahou and give it some more character understanding with Pantheon. Feedback is always welcome.

Native Context Length: 16K/16384 (can be extended using RoPE, YMMV)

Prompt Template: ChatML

<|im_start|>system
{system prompt}<|im_end|>
<|im_start|>user
{message}<|im_end|>
<|im_start|>assistant
{response}

Recommended Settings:

Here are some settings ranges that tend to work for me. They aren't strict values, and there's a bit of leeway in them. Feel free to experiment a bit!

• Temperature: 1.0 (maybe less, a little bit goes a long way with Nemo)
• Min-P: 0.1 to 0.2
• (all other samplers disabled)

Merge Details

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

Merge Method

This model was merged using the linear DARE merge method using flammenai/Mahou-1.3-mistral-nemo-12B as a base.

Models Merged

The following models were included in the merge:

• nbeerbower/mistral-nemo-gutenberg-12B-v4
• Sao10K/MN-12B-Lyra-v1
• Gryphe/Pantheon-RP-1.5-12b-Nemo
• flammenai/Mahou-1.3-mistral-nemo-12B

Configuration

The following YAML configuration was used to produce this model:

base_model: flammenai/Mahou-1.3-mistral-nemo-12B
dtype: bfloat16
merge_method: dare_linear
slices:
- sources:
  - layer_range: [0, 40]
    model: Gryphe/Pantheon-RP-1.5-12b-Nemo
    parameters:
      weight: [0.45, 0.35, 0.35, 0.2, 0.2]
  - layer_range: [0, 40]
    model: Sao10K/MN-12B-Lyra-v1
    parameters:
      weight: [0.25, 0.3, 0.35, 0.3, 0.2]
  - layer_range: [0, 40]
    model: nbeerbower/mistral-nemo-gutenberg-12B-v4
    parameters:
      weight:
      - filter: mlp
        value: [0.1, 0.2, 0.1, 0.4, 0.5]
      - value: [0.1, 0.2, 0.1, 0.2, 0.2]
  - layer_range: [0, 40]
    model: flammenai/Mahou-1.3-mistral-nemo-12B
    parameters:
      weight:
      - filter: mlp
        value: [0.2, 0.15, 0.2, 0.1, 0.1]
      - value: [0.2, 0.15, 0.2, 0.3, 0.4]
tokenizer_source: union

Benchmarks (or Benchmark because I tried only one)

I ran EQ bench from EleutherAI's lm-evaluation-harness (thank you @FallenMerick).

| Tasks  |Version|Filter|n-shot|     Metric      |   | Value  |   |Stderr|
|--------|------:|------|-----:|-----------------|---|-------:|---|-----:|
|eq_bench|    2.1|none  |     0|eqbench          |↑  | 79.3617|±  | 1.637|
|        |       |none  |     0|percent_parseable|↑  |100.0000|±  | 0.000|

And as always, have a great day!