Model Type | |
Use Cases |
Areas: | Research, Specialization, Finetuning |
|
Applications: | Summarization, Text generation, Chatbot |
|
Primary Use Cases: | Research on large language models, Foundation for further specialization |
|
Limitations: | Not suitable for production use without risk assessment and mitigation, Trained mostly on a few languages |
|
Considerations: | Finetuning is recommended for specific tasks. |
|
|
Additional Notes | Quantized version intended to be identical in licensing to the original huggingface model. |
|
Supported Languages | English (high proficiency), German (high proficiency), Spanish (high proficiency), French (high proficiency), Italian (limited capabilities), Portuguese (limited capabilities), Polish (limited capabilities), Dutch (limited capabilities), Romanian (limited capabilities), Czech (limited capabilities), Swedish (limited capabilities) |
|
Training Details |
Data Sources: | RefinedWeb, Books, Conversations, Code, Technical |
|
Data Volume: | |
Context Length: | |
Training Time: | |
Hardware Used: | |
Model Architecture: | Adapted from GPT-3 with rotary positionnal embeddings, multiquery and FlashAttention. |
|
|
Responsible Ai Considerations |
Fairness: | Falcon-40B carries stereotypes and biases present on the web. |
|
Transparency: | Details of datasets and methodology provided. |
|
Accountability: | It is recommended to further finetune the model and implement guardrails for production use. |
|
Mitigation Strategies: | Users are recommended to consider finetuning and appropriate precautions for production use. |
|
|
Input Output |
Input Format: | |
Accepted Modalities: | |
Output Format: | |
Performance Tips: | Use PyTorch 2.0 for optimal functionality. |
|
|
Release Notes |
Version: | |
Date: | |
Notes: | Converted using ct2-transformers-converter with int8_float16 quantization for improved efficiency. |
|
|
|