README.md

metadata

license: cc-by-nc-4.0
library_name: transformers
tags:
  - mergekit
  - merge
base_model:
  - grimjim/Mistral-Starling-merge-trial1-7B
  - grimjim/kukulemon-7B
pipeline_tag: text-generation
model-index:
  - name: cuckoo-starling-7B
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 66.81
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=grimjim/cuckoo-starling-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 85.97
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=grimjim/cuckoo-starling-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 64.88
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=grimjim/cuckoo-starling-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 59.03
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=grimjim/cuckoo-starling-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 80.11
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=grimjim/cuckoo-starling-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 62.77
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=grimjim/cuckoo-starling-7B
          name: Open LLM Leaderboard

cuckoo-starling-32k-7B

For this merged model, rope theta was in config.json was manually adjusted down to 100K, a value less than 1M as initially released by Mistral for v0.2, but higher than the 10K that accompanied practical 8K context for v0.1. We idly conjecture that 1M rope theta might improve performance for needle-in-a-haystack queries; however, during informal testing, narrative coherence seemed to occasionally suffer under 1M rope theta. Furthermore, the results reported in the arXiv paper Scaling Laws of RoPE-based Extrapolation suggest that 1M rope theta may be overkill for a 32K token context window.

Lightly tested with temperature 0.9-1.0 and minP 0.02, using ChatML prompts. The model natively supports Alpaca prompts.

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

• Full weights: grimjim/cuckoo-starling-32k-7B
• GGUFs: grimjim/cuckoo-starling-32k-7B-GGUF

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• grimjim/Mistral-Starling-merge-trial1-7B
• grimjim/kukulemon-7B

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
      - model: grimjim/Mistral-Starling-merge-trial1-7B
        layer_range: [0, 32]
      - model: grimjim/kukulemon-7B
        layer_range: [0, 32]
# or, the equivalent models: syntax:
# models:
merge_method: slerp
base_model: grimjim/Mistral-Starling-merge-trial1-7B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5 # fallback for rest of tensors
dtype: bfloat16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	69.93
AI2 Reasoning Challenge (25-Shot)	66.81
HellaSwag (10-Shot)	85.97
MMLU (5-Shot)	64.88
TruthfulQA (0-shot)	59.03
Winogrande (5-shot)	80.11
GSM8k (5-shot)	62.77