Yi 34B By 01-ai: Benchmarks, Features and Detailed Analysis. Insights on Yi 34B.

Arxiv:2311.16502 Arxiv:2401.11944 Arxiv:2403.04652 Autotrain compatible Endpoints compatible Llama Pytorch Region:us Safetensors Sharded Tensorflow

Yi 34B Benchmarks

MMLU Pro: 37.91

GPQA: 15.55

MUSR: 9.65

BBH: 35.54

IFEval: 30.46 vs 88 (so35)^-65.4%

ARC: 91.9 vs 96.7 (so35)^-5%

HellaSwag: 82 vs 95.3 (gpt4)^-14%

MMLU: 76.3 vs 88.3 (so35)^-13.6%

TruthfulQA: 56.23 vs 59 (gpt4)^-4.7%

WinoGrande: 83.03 vs 87.5 (gpt4)^-5.1%

GSM8K: 67.9 vs 96.4 (so35)^-29.6%

OpenCompas: 63.3 vs 73.3 (gpt4)^-13.6%

Exam: 78.5 vs 77.2 (gpt4)^1.7%

Language: 44.9 vs 62 (gpt4)^-27.6%

Knowledge: 64.5 vs 73.5 (gpt4)^-12.2%

Understand: 76.6 vs 70 (gpt4)^9.4%

Reasoning: 59 vs 74.4 (gpt4)^-20.7%

MATH Lvl 5: 5.21

LLME Score: 0.35738

^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)

Yi 34B Parameters and Internals

Model Type

Chat model, Text generation

Use Cases

Areas:

Chat applications, Creative content generation

Applications:

Commercial applications, Research, Educational tools

Primary Use Cases:

Chatbots, Virtual assistants, Story generation

Limitations:

Potential for hallucination, May produce inconsistent outputs

Considerations:

Adjust generation parameters for desired output qualities.

Additional Notes

Models do not directly use Llama's weights; unique datasets and training infrastructure emphasize Yi's independent development.

Supported Languages

English (Fluent), Chinese (Fluent)

Training Details

Data Sources:

Trainer Multilingual Corpora, 3T Tokens

Data Volume:

3T Multilingual Corpus

Methodology:

Transformer-based architecture

Context Length:

200000

Training Time:

Not specified

Hardware Used:

NVIDIA A800 (80GB), 4090 GPU

Model Architecture:

Based on Llama's architecture

Responsible Ai Considerations

Fairness:

Addressed during model development.

Transparency:

Standard Transformer architecture; detailed in tech report.

Accountability:

01.AI

Mitigation Strategies:

Use of Supervised Fine-Tuning for better accuracy.

Input Output

Input Format:

Interactive prompt conversation

Accepted Modalities:

Text

Output Format:

Text responses or follow-ups

Performance Tips:

Calibrate temperature, top_p, top_k settings for desired response diversity.

Release Notes

Version:

1.0

Date:

2023-11-23

Notes:

Initial open-source release of chat model, supporting both 4-bit and 8-bit quantizations.

Version:

2.0

Date:

2023-12-19

Notes:

Improved performance in coding, math, and reasoning with larger context capabilities.

LLM Name	Yi 34B
Repository 🤗	https://huggingface.co/01-ai/Yi-34B
Model Size	34b
Required VRAM	68.9 GB
Updated	2024-11-21
Maintainer	01-ai
Model Type	llama
Model Files	10.0 GB: 1-of-7 9.9 GB: 2-of-7 9.8 GB: 3-of-7 9.8 GB: 4-of-7 9.8 GB: 5-of-7 9.9 GB: 6-of-7 9.7 GB: 7-of-7 10.0 GB: 1-of-7 9.9 GB: 2-of-7 9.8 GB: 3-of-7 9.8 GB: 4-of-7 9.8 GB: 5-of-7 9.9 GB: 6-of-7 9.7 GB: 7-of-7
Model Architecture	LlamaForCausalLM
License	apache-2.0
Context Length	4096
Model Max Length	4096
Transformers Version	4.34.0
Tokenizer Class	LlamaTokenizer
Padding Token	<unk>
Vocabulary Size	64000
Torch Data Type	bfloat16

Quantized Models of the Yi 34B

Model	Likes	Downloads	VRAM
Yi 34B GGUF	76	631	14 GB
...Hermes 2 Yi 34B 4.0bpw H6 EXL2	4	2	18 GB
...Hermes 2 Yi 34B 3.0bpw H6 EXL2	2	5	13 GB
...Hermes 2 Yi 34B 5.0bpw H6 EXL2	1	3	22 GB
...ermes 2 Yi 34B 4.65bpw H6 EXL2	1	3	20 GB
Yi 34B GPTQ	33	65	18 GB
Yi 34B AWQ	7	12	19 GB

Best Alternatives to Yi 34B

Best Alternatives	Context / RAM	Downloads	Likes
Bagel 34B V0.2	195K / 68.7 GB	15531	39
34B Beta	195K / 69.2 GB	3730	62
Yi 34B 200K	195K / 68.9 GB	4698	316
Bagel Hermes 34B Slerp	195K / 68.9 GB	3977	1
Smaug 34B V0.1	195K / 69.2 GB	3201	60
Yi 34B 200K AEZAKMI V2	195K / 69.2 GB	1254	12
Mergekit Slerp Anaazls	195K / 69.2 GB	37	0
Smaug 34B V0.1 ExPO	195K / 69.2 GB	2844	0
Faro Yi 34B	195K / 69.2 GB	3637	6
Bagel DPO 34B V0.5	195K / 68.7 GB	2873	17

Rank the Yi 34B 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? 38149 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 v20241110

Support LLM Explorer