Model Type | |
Use Cases |
Areas: | chat applications, instruct models |
|
Applications: | virtual assistants, content generation, customer support |
|
Primary Use Cases: | chat responses, text generation |
|
Limitations: | Not suitable for non-European languages, May contain biases inherent from web data |
|
Considerations: | Implement appropriate guardrails and precautions for production use. |
|
|
Additional Notes | Powered by NTK-YaRN, improving long context performance. |
|
Supported Languages | English (high), French (high), German (high), Spanish (high), Italian (limited), Portuguese (limited), Polish (limited), Dutch (limited), Romanian (limited), Czech (limited), Swedish (limited) |
|
Training Details |
Data Sources: | OpenAssistant/oasst1, ehartford/dolphin, openai-critiques, tau/sled, internal, internal-long-context, RefinedWeb |
|
Data Volume: | |
Methodology: | Supervised finetuning; context length extension with NTK-YaRN; training on short and long context tasks |
|
Context Length: | |
Hardware Used: | |
Model Architecture: | Causal decoder; extended context length of 8192 tokens using NTK-YaRN |
|
|
Responsible Ai Considerations |
Fairness: | Training on diverse languages though primarily on English, German, Spanish and French; limited language support for others. |
|
|
Input Output |
Input Format: | Text prompts, possibly with chat tokens for chat tasks. |
|
Accepted Modalities: | |
Output Format: | |
Performance Tips: | Use chat token formats for chat tasks to guide behavior. |
|
|
Release Notes |
Version: | |
Date: | |
Notes: | Initial release of Alfred 40B 1023 with 8K token context extension. |
|
|
|