Model Card: Hercules-4.0-Mistral-v0.2-7B

Model Description

Hercules-4.0-Mistral-v0.2-7B is a fine-tuned language model derived from Mistralai/Mistral-7B-v0.2. It is specifically designed to excel in instruction following, function calls, and conversational interactions across various scientific and technical domains. The dataset used for fine-tuning, also named Hercules-v4.0, expands upon the diverse capabilities of OpenHermes-2.5 with contributions from numerous curated datasets. This fine-tuning has hercules-v4.0 with enhanced abilities in:

• Complex Instruction Following: Understanding and accurately executing multi-step instructions, even those involving specialized terminology.
• Function Calling: Seamlessly interpreting and executing function calls, providing appropriate input and output values.
• Domain-Specific Knowledge: Engaging in informative and educational conversations about Biology, Chemistry, Physics, Mathematics, Medicine, Computer Science, and more.

Quants

GGUF: https://huggingface.co/bartowski/Hercules-4.0-Mistral-v0.2-7B-GGUF

ExLlamaV2: https://huggingface.co/bartowski/Hercules-4.0-Mistral-v0.2-7B-exl2

AWQ: https://huggingface.co/solidrust/Hercules-4.0-Mistral-v0.2-7B-AWQ

Intended Uses & Potential Bias

Hercules-4.0-Mistral-v.02-7B is well-suited to the following applications:

• Specialized Chatbots: Creating knowledgeable chatbots and conversational agents in scientific and technical fields.
• Instructional Assistants: Supporting users with educational and step-by-step guidance in various disciplines.
• Code Generation and Execution: Facilitating code execution through function calls, aiding in software development and prototyping.

Important Note: Although Hercules-v4.0 is carefully constructed, it's important to be aware that the underlying data sources may contain biases or reflect harmful stereotypes. Use this model with caution and consider additional measures to mitigate potential biases in its responses.

Limitations and Risks

• Toxicity: The dataset contains toxic or harmful examples.
• Hallucinations and Factual Errors: Like other language models, Hercules-4.0-Mistral-v0.2-7B may generate incorrect or misleading information, especially in specialized domains where it lacks sufficient expertise.
• Potential for Misuse: The ability to engage in technical conversations and execute function calls could be misused for malicious purposes.

Training Procedure

• This model was trained on 8 kaggle TPUs, using torch xla SPMD for high MXU efficiency. There was no expense on my end (meaning you can reproduce this too!)
• A learning rate of 5e-06 with the Adam optimizer. A linear scheduler was used, with an end factor of 0.1. A low learning rate was used to prevent exploding gradients.
• No mixed precision was used, with the default dtype being bfloat16.
• A total batch size of 64 was used.
• Trained on 700,000 examples of Hercules-v4.0
• No model parameters were frozen and no quantization was used.
• This model was trained on OpenAI's ChatML prompt format. Because this model has function calling capabilities, the prompt format is slightly different, here's what it would look like: <|im_start|>system\n{message}<|im_end|>\n<|im_start|>user\n{user message}<|im_end|>\n<|im_start|>call\n{function call message}<|im_end|>\n<|im_start|>function\n{function response message}<|im_end|>\n<|im_start|>assistant\n{assistant message}</s>

This model was fine-tuned using my TPU-Alignment repository. https://github.com/Locutusque/TPU-Alignment

Evaluation

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
agieval_nous	N/A	none	0	acc_norm	0.3578	±	0.0093
		none	0	acc	0.3598	±	0.0093
- agieval_aqua_rat	1	none	0	acc	0.2362	±	0.0267
		none	0	acc_norm	0.2362	±	0.0267
- agieval_logiqa_en	1	none	0	acc	0.3026	±	0.0180
		none	0	acc_norm	0.3456	±	0.0187
- agieval_lsat_ar	1	none	0	acc	0.2565	±	0.0289
		none	0	acc_norm	0.2522	±	0.0287
- agieval_lsat_lr	1	none	0	acc	0.3588	±	0.0213
		none	0	acc_norm	0.3490	±	0.0211
- agieval_lsat_rc	1	none	0	acc	0.4981	±	0.0305
		none	0	acc_norm	0.4461	±	0.0304
- agieval_sat_en	1	none	0	acc	0.6553	±	0.0332
		none	0	acc_norm	0.6408	±	0.0335
- agieval_sat_en_without_passage	1	none	0	acc	0.3883	±	0.0340
		none	0	acc_norm	0.3738	±	0.0338
- agieval_sat_math	1	none	0	acc	0.3091	±	0.0312
		none	0	acc_norm	0.2773	±	0.0302

Groups	Version	Filter	n-shot	Metric	Value		Stderr
agieval_nous	N/A	none	0	acc_norm	0.3578	±	0.0093
		none	0	acc	0.3598	±	0.0093
mmlu_flan_cot_fewshot	N/A	get-answer	0	exact_match	0.5787	±	0.0120
- mmlu_flan_cot_fewshot_humanities	N/A	get-answer	0	exact_match	0.5174	±	0.0209
- mmlu_flan_cot_fewshot_other	N/A	get-answer	0	exact_match	0.6833	±	0.0249
- mmlu_flan_cot_fewshot_social_sciences	N/A	get-answer	0	exact_match	0.6528	±	0.0248
- mmlu_flan_cot_fewshot_stem	N/A	get-answer	0	exact_match	0.4925	±	0.0266
ai2_arc	N/A	none	0	acc	0.6936	±	0.0073
		none	0	acc_norm	0.6984	±	0.0074

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	61.53
AI2 Reasoning Challenge (25-Shot)	58.96
HellaSwag (10-Shot)	82.60
MMLU (5-Shot)	62.66
TruthfulQA (0-shot)	40.99
Winogrande (5-shot)	78.53
GSM8k (5-shot)	45.41