Breeze 7B Instruct V1.0 By MediaTek-Research: Benchmarks, Features and Detailed Analysis. Insights on Breeze 7B Instruct V1.0.

Arxiv:2403.02712 Autotrain compatible Conversational En Endpoints compatible Instruct Mistral Pytorch Region:us Safetensors Sharded Tensorflow Zh

Model Card on HF 🤗: https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v1_0

Breeze 7B Instruct V1.0 Benchmarks

ARC: 61.35 vs 96.7 (so35)^-36.6%

HellaSwag: 82.59 vs 95.3 (gpt4)^-13.3%

MMLU: 62.11 vs 88.3 (so35)^-29.7%

TruthfulQA: 53.46 vs 59 (gpt4)^-9.4%

WinoGrande: 77.74 vs 87.5 (gpt4)^-11.2%

GSM8K: 43.14 vs 96.4 (so35)^-55.2%

LLME Score: 0.25321

^nn.n% — How the model compares to the reference models: Anthropic Sonnet 3.5 ("so35"), GPT-4o ("gpt4o") or GPT-4 ("gpt4").

What is the LLM Explorer Rank (Score)

Breeze 7B Instruct V1.0 (MediaTek-Research/Breeze-7B-Instruct-v1_0)

Breeze 7B Instruct V1.0 Parameters and Internals

Model Type

Causal decoder-only transformer

Use Cases

Areas:

Research, Commercial applications

Applications:

Q&A, RAG, Multi-round chat, Summarization

Primary Use Cases:

Commonly seen tasks

Limitations:

No special handling for harmfulness

Additional Notes

Expanded vocabulary with 30,000 Traditional Chinese tokens. Faster inference for Traditional Chinese compared to other models.

Supported Languages

English (Supported), Traditional Chinese (Supported)

Training Details

Methodology:

Fine-tuning from base model.

Context Length:

8000

Model Architecture:

Causal decoder-only transformer

Input Output

Input Format:

~~SYS_PROMPT [INST] QUERY1 [/INST] RESPONSE1 [INST] QUERY2 [/INST]

Accepted Modalities:

text

Output Format:

text

LLM Name	Breeze 7B Instruct V1.0
Repository 🤗	https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v1_0
Model Size	7b
Required VRAM	15 GB
Updated	2024-12-21
Maintainer	MediaTek-Research
Model Type	mistral
Instruction-Based	Yes
Model Files	5.0 GB: 1-of-4 4.9 GB: 2-of-4 4.6 GB: 3-of-4 0.5 GB: 4-of-4
Supported Languages	zh en
Model Architecture	MistralForCausalLM
License	apache-2.0
Context Length	32768
Model Max Length	32768
Transformers Version	4.37.2
Tokenizer Class	LlamaTokenizer
Padding Token	</s>
Vocabulary Size	61952
Torch Data Type	bfloat16

Quantized Models of the Breeze 7B Instruct V1.0

Model	Likes	Downloads	VRAM
Breeze 7B Instruct V1.0 AWQ	0	25	4 GB
...reeze 7B 32K Instruct V1.0 AWQ	0	9	4 GB
Breeze 7B Instruct V1.0 AWQ	1	18	4 GB
...Instruct V1.0 Bnb 4bit Smashed	0	41	4 GB

Best Alternatives to Breeze 7B Instruct V1.0

Best Alternatives	Context / RAM	Downloads	Likes
...Nemo Instruct 2407 Abliterated	1000K / 24.5 GB	2668	9
SpydazWeb AI HumanAI RP	512K / 14.4 GB	33	1
SpydazWeb AI HumanAI 002	512K / 14.4 GB	19	1
...daz Web AI ChatML 512K Project	512K / 14.5 GB	12	0
... Summarize 64K QLoRANET Merged	128K / 4.1 GB	7	0
...1 Summarize 64K LoRANET Merged	128K / 14.4 GB	6	0
Mistral 7B Instruct V0.2	32K / 14.4 GB	3174759	2597
Mistral 7B Instruct V0.1	32K / 14.4 GB	3139801	1538
...ity Instruct 7M Gen Mistral 7B	32K / 14.4 GB	5406	4
...ty Instruct 3M 0625 Mistral 7B	32K / 14.4 GB	5420	3

Note: green Score (e.g. "73.2") means that the model is better than MediaTek-Research/Breeze-7B-Instruct-v1_0.

Rank the Breeze 7B Instruct V1.0 Capabilities

🆘 Have you tried this model? Rate its performance. This feedback would greatly assist ML community in identifying the most suitable model for their needs. Your contribution really does make a difference! 🌟

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

What open-source LLMs or SLMs are you in search of? 40066 in total.

Email us: info@extractum.io. Our Privacy Policy | Terms and Conditions | Suggest an improvement.

Our Social Media →

Original data from HuggingFace, OpenCompass and various public git repos.

Release v20241217

Support LLM Explorer