Llama-3-NeuralHercules-5.0-8B

Model Details

• Model Name: Locutusque/Llama-3-NeuralHercules-5.0-8B
• Base Model: meta-llama/Meta-Llama-3-8B
• Publisher: Locutusque
• Model Type: Question answering, conversational AI, code generation, medical text comprehension, mathematical reasoning, logical reasoning.
• Language: multi-lingual
• License: Apache-2.0

Model Description

Locutusque/Llama-3-NeuralHercules-5.0-8B is a state-of-the-art language model fine-tuned on the hercules-v5.0 and mlabonne/orpo-dpo-mix-40k dataset for advanced reasoning across scientific domains. This model is designed to handle complex inquiries and instructions, leveraging the diverse and rich information contained in the Hercules dataset. Its primary use cases include but are not limited to complex question answering, conversational understanding, code generation, medical text comprehension, mathematical reasoning, and logical reasoning.

Intended Use

This model is intended for researchers and practitioners looking for a powerful tool to tackle challenging problems in scientific domains. It can be used in the following scenarios:

• AI-driven tutoring systems for science, medicine, mathematics, and computer science.
• Assistive tools for professionals requiring fast and accurate domain-specific information retrieval.
• Platforms that require conversational AI capabilities with a focus on technical and scientific reasoning.
• Automation in code generation and understanding complex programming context.

Training Data

The Locutusque/Llama-3-NeuralHercules-5.0-8B model was fine-tuned on all examples of the hercules-v5.0 dataset for 2 epochs, which amalgamates various datasets rich in diversity and complexity, including programming, medical texts, mathematical problems, and reasoning tasks. Then, it is further fine-tuned on the mlabonne/orpo-dpo-mix-40k dataset using DPO.

Evaluation Results

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
agieval_nous	N/A	none	0	acc_norm	↑	0.3943	±	0.0094
		none	0	acc	↑	0.3928	±	0.0094
- agieval_aqua_rat	1	none	0	acc	↑	0.2126	±	0.0257
		none	0	acc_norm	↑	0.2323	±	0.0265
- agieval_logiqa_en	1	none	0	acc	↑	0.3425	±	0.0186
		none	0	acc_norm	↑	0.3625	±	0.0189
- agieval_lsat_ar	1	none	0	acc	↑	0.2174	±	0.0273
		none	0	acc_norm	↑	0.2000	±	0.0264
- agieval_lsat_lr	1	none	0	acc	↑	0.4020	±	0.0217
		none	0	acc_norm	↑	0.3843	±	0.0216
- agieval_lsat_rc	1	none	0	acc	↑	0.5651	±	0.0303
		none	0	acc_norm	↑	0.5651	±	0.0303
- agieval_sat_en	1	none	0	acc	↑	0.6602	±	0.0331
		none	0	acc_norm	↑	0.6456	±	0.0334
- agieval_sat_en_without_passage	1	none	0	acc	↑	0.4466	±	0.0347
		none	0	acc_norm	↑	0.4563	±	0.0348
- agieval_sat_math	1	none	0	acc	↑	0.4000	±	0.0331
		none	0	acc_norm	↑	0.4000	±	0.0331
gsm8k	3	strict-match	5	exact_match	↑	0.4920	±	0.0138
		flexible-extract	5	exact_match	↑	0.4958	±	0.0138
truthfulqa_mc2	2	none	0	acc	↑	0.5465	±	0.0152
arc_challenge	1.0	none	0	acc	↑	0.5606	±	0.0145
		none	0	acc_norm	↑	0.5836	±	0.0144
arc_easy	1	none	0	acc	↑	0.8325	±	0.0077
		none	0	acc_norm	↑	0.8056	±	0.0081
boolq	2	none	0	acc	↑	0.8260	±	0.0066
hellaswag	1	none	0	acc	↑	0.6534	±	0.0047
		none	0	acc_norm	↑	0.8372	±	0.0037
openbookqa	1	none	0	acc	↑	0.3500	±	0.0214
		none	0	acc_norm	↑	0.4660	±	0.0223
piqa	1	none	0	acc	↑	0.8096	±	0.0092
		none	0	acc_norm	↑	0.8270	±	0.0088
winogrande	1	none	0	acc	↑	0.7640	±	0.0119
eq_bench	2.1	none	0	eqbench	↑	52.9249	±	3.0923

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Locutusque/Llama-3-NeuralHercules-5.0-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# For a text generation task
input_text = "<|im_start|>user\nWhat are the implications of Einstein's theory of relativity in modern physics?<|im_end|>\n<|im_start|>assistant\n"
input_ids = tokenizer.encode(input_text, return_tensors="pt")

# Generate a response
outputs = model.generate(input_ids, max_length=200, num_return_sequences=1, temperature=0.8, top_p=0.95, top_k=40, repetition_penalty=1.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Known Limitations

The diversity of the dataset could lead to inconsistencies in the model's responses due to variations in data formatting and annotation quality.