Marcoro14-7B-slerp

This model is a merge of the following models made with mergekit:

🏆 Evaluation

Marcoro14-7B-slerp is the best-performing 7B LLM on the Open LLM Leaderboard (rank 1 below is 9B):

I also evaluated it using Nous' benchmark suite and obtained the following results:

Model	AGIEval	GPT4ALL	TruthfulQA	Bigbench	Average
Marcoro14-7B-slerp	44.66	76.24	64.15	45.64	57.67
OpenHermes-2.5-Mistral-7B	43.07	73.12	53.04	40.96	52.57
Change	+1.59	+3.12	+11.11	+4.68	+5.1

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	26.38	±	2.77
		acc_norm	24.41	±	2.70
agieval_logiqa_en	0	acc	38.25	±	1.91
		acc_norm	39.32	±	1.92
agieval_lsat_ar	0	acc	24.35	±	2.84
		acc_norm	25.22	±	2.87
agieval_lsat_lr	0	acc	50.00	±	2.22
		acc_norm	50.59	±	2.22
agieval_lsat_rc	0	acc	62.83	±	2.95
		acc_norm	62.08	±	2.96
agieval_sat_en	0	acc	79.61	±	2.81
		acc_norm	79.61	±	2.81
agieval_sat_en_without_passage	0	acc	45.15	±	3.48
		acc_norm	45.63	±	3.48
agieval_sat_math	0	acc	33.18	±	3.18
		acc_norm	30.45	±	3.11

Average: 44.66%

GPT4ALL

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	63.91	±	1.40
		acc_norm	64.93	±	1.39
arc_easy	0	acc	86.07	±	0.71
		acc_norm	83.75	±	0.76
boolq	1	acc	88.56	±	0.56
hellaswag	0	acc	67.31	±	0.47
		acc_norm	85.28	±	0.35
openbookqa	0	acc	36.40	±	2.15
		acc_norm	48.20	±	2.24
piqa	0	acc	82.59	±	0.88
		acc_norm	84.39	±	0.85
winogrande	0	acc	78.53	±	1.15

Average: 76.24%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	46.88	±	1.75
		mc2	64.15	±	1.52

Average: 64.15%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	56.32	±	3.61
bigbench_date_understanding	0	multiple_choice_grade	66.40	±	2.46
bigbench_disambiguation_qa	0	multiple_choice_grade	45.35	±	3.11
bigbench_geometric_shapes	0	multiple_choice_grade	20.33	±	2.13
		exact_str_match	4.74	±	1.12
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	30.00	±	2.05
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	21.43	±	1.55
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	52.33	±	2.89
bigbench_movie_recommendation	0	multiple_choice_grade	39.20	±	2.19
bigbench_navigate	0	multiple_choice_grade	53.90	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	72.15	±	1.00
bigbench_ruin_names	0	multiple_choice_grade	52.46	±	2.36
bigbench_salient_translation_error_detection	0	multiple_choice_grade	25.75	±	1.38
bigbench_snarks	0	multiple_choice_grade	72.38	±	3.33
bigbench_sports_understanding	0	multiple_choice_grade	73.63	±	1.40
bigbench_temporal_sequences	0	multiple_choice_grade	45.70	±	1.58
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	23.44	±	1.20
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	18.51	±	0.93
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	52.33	±	2.89

Average: 45.64%

Average score: 57.67%

🧩 Configuration

slices:
  - sources:
      - model: AIDC-ai-business/Marcoroni-7B-v3
        layer_range: [0, 32]
      - model: EmbeddedLLM/Mistral-7B-Merge-14-v0.1
        layer_range: [0, 32]
merge_method: slerp
base_model: AIDC-ai-business/Marcoroni-7B-v3
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
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/Marcoro14-7B-slerp"
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"])

Output:

A large language model is a type of artificial intelligence (AI) system that has been trained on vast amounts of text data. It's designed to understand and generate human-like language, making predictions on what words or phrases might come next in a sentence or document. These models use complex algorithms and neural network architectures to learn from the data and improve their performance over time. Some well-known large language models include GPT-3 from OpenAI and BERT from Google.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	73.01
AI2 Reasoning Challenge (25-Shot)	69.80
HellaSwag (10-Shot)	87.13
MMLU (5-Shot)	65.11
TruthfulQA (0-shot)	63.54
Winogrande (5-shot)	81.61
GSM8k (5-shot)	70.89

mlabonne
/

Marcoro14-7B-slerp