Llama3-8B-SuperNova-Spectrum-dare_ties

Llama3-8B-SuperNova-Spectrum-dare_ties is a dare_ties merge of the following models using LazyMergekit:

• yuvraj17/Llama-3-8B-spectrum-25
• ruggsea/Llama3-stanford-encyclopedia-philosophy-QA
• arcee-ai/Llama-3.1-SuperNova-Lite

DARE_TIES Merging

TIES Merging

TIES Merging, introduced by Yadav et al. (2023), is a method for merging multiple specialized models into one general-purpose model. It solves two key challenges:

• Redundancy Removal: Identifies and eliminates overlapping or unnecessary information between models, making the final model more efficient.
• Conflict Resolution: Reconciles differences between models by creating a unified sign vector that represents the most dominant direction of change across all models.

TIES stands for TRIM, ELECT SIGN & MERGE (TIES-MERGING).

How TIES-Merging Works Reference

DARE Merging

Introduced by Yu et al. (2023), DARE uses an approach similar to TIES with two main differences:

• Weight Pruning: Randomly resets some fine-tuned weights to their original values, reducing model complexity.
• Weight Scaling: Adjusts the remaining weights by scaling and combining them with the base model's weights to maintain consistent performance.

DARE stands for DROP AND RESCALE

Mergekit’s implementation of DARE-Merging has two flavours: with the sign election step of TIES (dare_ties) or without (dare_linear). I have chosen dare_ties for this merge.

For more information refer this Merge Large Language Models with MergeKit by Maxime Labonne

Also, if you want to get in-depth knowledge about Model-Merging and its different types, I highly recommend this YouTube Video by Julien Simon

🧩 Configuration

models:
  - model: NousResearch/Meta-Llama-3-8B
    # No parameters necessary for base model
  - model: yuvraj17/Llama-3-8B-spectrum-25
    parameters:
      density: 0.56
      weight: 0.12
  - model: ruggsea/Llama3-stanford-encyclopedia-philosophy-QA
    parameters:
      density: 0.56
      weight: 0.12
  - model: arcee-ai/Llama-3.1-SuperNova-Lite
    parameters:
      density: 0.58
      weight: 0.55
merge_method: dare_ties
base_model: NousResearch/Meta-Llama-3-8B
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "yuvraj17/Llama3-8B-SuperNova-Spectrum-dare_ties"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

A large language model is a type of artificial intelligence (AI) model designed to understand and generate human language. It is trained on a massive corpus of text data, which it uses to learn patterns and relationships between words and concepts. Large language models are typically based on a deep learning approach called transformer architecture, which was introduced by the Google research paper "Attention Is All You Need" (2017). These models are designed to handle the complexity of natural language by capturing long-range dependencies and contextual relationships between words. Large language models can perform a variety of tasks, including:

• Natural language processing (NLP): large language models can understand and generate text, and can be used for tasks such as text classification, sentiment analysis, and named entity recognition.
• Text generation: large language models can generate human-like text, such as chatbots, language translation, and text summarization.
• Question answering: large language models can answer questions based on the text they have been trained on.
• Conversational AI: large language models can be used to create conversational agents that can understand and respond to user input.

🏆 Evaluation Scores

Nous

Model	AGIEval	TruthfulQA	Bigbench
Llama3-8B-SuperNova-Spectrum-dare_ties	38.32	57.15	43.91

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	20.47	±	2.54
		acc_norm	18.50	±	2.44
agieval_logiqa_en	0	acc	35.94	±	1.88
		acc_norm	35.64	±	1.88
agieval_lsat_ar	0	acc	21.74	±	2.73
		acc_norm	20.00	±	2.64
agieval_lsat_lr	0	acc	41.37	±	2.18
		acc_norm	40.98	±	2.18
agieval_lsat_rc	0	acc	59.11	±	3.00
		acc_norm	56.13	±	3.03
agieval_sat_en	0	acc	63.59	±	3.36
		acc_norm	60.19	±	3.42
agieval_sat_en_without_passage	0	acc	40.29	±	3.43
		acc_norm	37.38	±	3.38
agieval_sat_math	0	acc	38.64	±	3.29
		acc_norm	37.73	±	3.28

Average: 38.32%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	38.43	±	1.7
		mc2	57.15	±	1.5

Average: 57.15%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	58.42	±	3.59
bigbench_date_understanding	0	multiple_choice_grade	70.73	±	2.37
bigbench_disambiguation_qa	0	multiple_choice_grade	30.23	±	2.86
bigbench_geometric_shapes	0	multiple_choice_grade	47.35	±	2.64
		exact_str_match	0.00	±	0.00
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	29.00	±	2.03
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	21.00	±	1.54
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	51.33	±	2.89
bigbench_movie_recommendation	0	multiple_choice_grade	33.20	±	2.11
bigbench_navigate	0	multiple_choice_grade	55.40	±	1.57
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	66.35	±	1.06
bigbench_ruin_names	0	multiple_choice_grade	45.76	±	2.36
bigbench_salient_translation_error_detection	0	multiple_choice_grade	28.26	±	1.43
bigbench_snarks	0	multiple_choice_grade	62.43	±	3.61
bigbench_sports_understanding	0	multiple_choice_grade	50.30	±	1.59
bigbench_temporal_sequences	0	multiple_choice_grade	48.00	±	1.58
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	23.60	±	1.20
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	17.66	±	0.91
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	51.33	±	2.89

Average: 43.91%

Special thanks & Reference

• Maxime Labonne for their easy-to-use colab-notebook Merging LLMs with MergeKit, Blog and LLM-AutoEva Notebookl
• Authors of Mergekit

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	19.00
IFEval (0-Shot)	40.13
BBH (3-Shot)	23.49
MATH Lvl 5 (4-Shot)	7.40
GPQA (0-shot)	3.36
MuSR (0-shot)	11.00
MMLU-PRO (5-shot)	28.60