Shining Valiant is a chat model built on the Llama 2 architecture, finetuned on our data for insight, creativity, passion, and friendliness.
- Uses the llama-2-70b-chat model, with safetensors
- Finetuned on multiple runs across private and public data
- Data focused on knowledge, enthusiasm, and structured reasoning
Version
The current version is 1.2; congrats to our team on the new release!
Previous versions remain available in the repository. New models will be released for everyone once our team's training and validation process is complete :)
Evaluation
| Model | Avg | ARC | HS | MMLU | TQA |
|---|---|---|---|---|---|
| Shining Valiant 1.2 | 74.17 | 72.95 | 87.88 | 70.97 | 64.88 |
| Llama 2 | 67.35 | 67.32 | 87.33 | 69.83 | 44.92 |
| Llama 2 Chat | 66.80 | 64.59 | 85.88 | 63.91 | 52.80 |
Prompting Guide
Shining Valiant uses the same prompt format as Llama 2 Chat - feel free to use your existing prompts and scripts! A few examples of different formats:
[INST] Good morning! Can you let me know how to parse a text file and turn the semicolons into commas? [/INST]
[INST] (You are an intelligent, helpful AI assistant.) Hello, can you write me a thank you letter? [/INST]
[INST] << SYS >>You are an intelligent, helpful AI assistant.<< /SYS >>Deep dive about a country with interesting history: [/INST]
The Model
Shining Valiant is built on top of Stellar Bright, which uses Llama 2's 70b parameter architecture and features upgraded general capability. (Stellar Bright uses public open source data only.)
From there, we've created Shining Valiant through multiple finetuning runs on different compositions of our private dataset.
Our private data focuses primarily on applying Shining Valiant's personality: she's friendly, enthusiastic, insightful, knowledgeable, and loves to learn!
We are actively working on expanding and improving the Shining Valiant dataset for use in future releases of this model and others.
Shining Valiant is created by Valiant Labs. We care about open source. For everyone to use.
We encourage others to finetune further from our models.
- Downloads last month
- 17