README.md

metadata

language:
  - en
license: apache-2.0
datasets:
  - ehartford/dolphin
  - jondurbin/airoboros-2.2.1
  - ehartford/dolphin-coder
  - teknium/openhermes
  - ise-uiuc/Magicoder-OSS-Instruct-75K
  - ise-uiuc/Magicoder-Evol-Instruct-110K
  - LDJnr/Capybara
model-index:
  - name: dolphin-2.6-mistral-7b-dpo-5.93B
    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: 38.99
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Mihaiii/dolphin-2.6-mistral-7b-dpo-5.93B
          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: 61.01
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Mihaiii/dolphin-2.6-mistral-7b-dpo-5.93B
          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: 27.32
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Mihaiii/dolphin-2.6-mistral-7b-dpo-5.93B
          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: 53.51
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Mihaiii/dolphin-2.6-mistral-7b-dpo-5.93B
          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: 62.67
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Mihaiii/dolphin-2.6-mistral-7b-dpo-5.93B
          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: 0.23
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Mihaiii/dolphin-2.6-mistral-7b-dpo-5.93B
          name: Open LLM Leaderboard

This is a pruned version of cognitivecomputations/dolphin-2.6-mistral-7b-dpo from 7.24B params to 5.93B params (~ 82%).

Steps to replicate:

Use laserQlora.ipynb from cognitivecomputations/laserRMT to determine which layers should be eliminated.

Replace model_name = "mistralai/Mistral-7B-v0.1" with model_name = "cognitivecomputations/dolphin-2.6-mistral-7b-dpo". I also ran the script only for self_attn.v_proj (so change the script to layer_types=["self_attn.v_proj"])

Order by snr descending and eliminate top layers using mergekit. The threshold for elimination is up to you, depeding on how many layers you want removed. I decided to remove 6 layers (indexes: 3, 5, 16, 18, 19, 24 )

Here is the mergekit config:

slices:
  - sources:
    - model: "cognitivecomputations/dolphin-2.6-mistral-7b-dpo"
      layer_range: [0, 3]
  - sources:
    - model: "cognitivecomputations/dolphin-2.6-mistral-7b-dpo"
      layer_range: [4, 5]
  - sources:
    - model: "cognitivecomputations/dolphin-2.6-mistral-7b-dpo"
      layer_range: [6, 16]
  - sources:
    - model: "cognitivecomputations/dolphin-2.6-mistral-7b-dpo"
      layer_range: [17, 18]
  - sources:
    - model: "cognitivecomputations/dolphin-2.6-mistral-7b-dpo"
      layer_range: [20, 24]
  - sources:
    - model: "cognitivecomputations/dolphin-2.6-mistral-7b-dpo"
      layer_range: [25, 32]
merge_method: passthrough
dtype: bfloat16

The model outputted by mergekit with this configuration is this model (dolphin-2.6-mistral-7b-dpo-5.93B).

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	40.62
AI2 Reasoning Challenge (25-Shot)	38.99
HellaSwag (10-Shot)	61.01
MMLU (5-Shot)	27.32
TruthfulQA (0-shot)	53.51
Winogrande (5-shot)	62.67
GSM8k (5-shot)	0.23