README.md

metadata

license: apache-2.0
tags:
  - mergekit
  - merge
  - solar
base_model:
  - upstage/SOLAR-10.7B-Instruct-v1.0
  - NousResearch/Nous-Hermes-2-SOLAR-10.7B
model-index:
  - name: franken-SOLAR-18B-v1.0
    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: 65.53
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vicgalle/franken-SOLAR-18B-v1.0
          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: 86.45
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vicgalle/franken-SOLAR-18B-v1.0
          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: 63.72
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vicgalle/franken-SOLAR-18B-v1.0
          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: 62.14
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vicgalle/franken-SOLAR-18B-v1.0
          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: 78.53
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vicgalle/franken-SOLAR-18B-v1.0
          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: 45.79
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vicgalle/franken-SOLAR-18B-v1.0
          name: Open LLM Leaderboard

vicgalle/franken-SOLAR-18B-v1.0

This is a SOLAR-like model upscaled to 18B. It is a frankenmerge model created using mergekit, alternating layers of Nous-Hermes-2-SOLAR-10.7B and SOLAR-10.7B-Instruct.

Evaluations coming soon!

This model has very good writing capabilities (compared to SOLAR-10.7B), specially for role-playing.

Quantized GGUF variants here https://huggingface.co/vicgalle/franken-SOLAR-18B-v1.0-GGUF

Merge Details

Merge Method

This model was merged using the passthrough merge method.

Models Merged

The following models were included in the merge:

• upstage/SOLAR-10.7B-Instruct-v1.0
• NousResearch/Nous-Hermes-2-SOLAR-10.7B

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
    - model: NousResearch/Nous-Hermes-2-SOLAR-10.7B
      layer_range: [0, 12]
  - sources:
    - model: upstage/SOLAR-10.7B-Instruct-v1.0
      layer_range: [6, 18]
  - sources:
    - model: NousResearch/Nous-Hermes-2-SOLAR-10.7B
      layer_range: [13, 25]
  - sources:
    - model: upstage/SOLAR-10.7B-Instruct-v1.0
      layer_range: [19, 31]
  - sources:
    - model: NousResearch/Nous-Hermes-2-SOLAR-10.7B
      layer_range: [26, 38]
  - sources:
    - model: upstage/SOLAR-10.7B-Instruct-v1.0
      layer_range: [32, 44]
  - sources:
    - model: NousResearch/Nous-Hermes-2-SOLAR-10.7B
      layer_range: [39, 48]
    
merge_method: passthrough
dtype: float16

Usage

You can use the provided template:

tokenizer = AutoTokenizer.from_pretrained("vicgalle/franken-SOLAR-18B-v1.0")
model = AutoModelForCausalLM.from_pretrained("vicgalle/franken-SOLAR-18B-v1.0", torch_dtype=torch.float16, load_in_4bit=True)

conversation = [ {'role': 'system', 'content': SYSTEM_PROMPT}, {'role': 'user', 'content': USER_PROMPT} ] 
prompt = tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, use_cache=True, max_new_tokens=1024, do_sample=True, temperature=0.8)
output_text = tokenizer.decode(outputs[0]) 

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	67.03
AI2 Reasoning Challenge (25-Shot)	65.53
HellaSwag (10-Shot)	86.45
MMLU (5-Shot)	63.72
TruthfulQA (0-shot)	62.14
Winogrande (5-shot)	78.53
GSM8k (5-shot)	45.79