Even with the new GPT4.o there are broken links for ChatRTX developer documentation. But even when they work, often the information is dated and misleading. I understand, it is great that such rapid progress is happening with Tensor-Rag features. But may I suggest the following text be added to the
RAG\DataSet\ReadMe.md file (if my understanding is incorrect, please edit):
"Place your text data here.
Note that as an alpha product most of the documentation online is out of date and can be misleading.
(Warning: dataset txt rules are not the same for different llms, and are evolving all the time. e.g. Mistral vs. Lama)
The Nvidia Tensor-RAG goal is for the dataset syntax \ punctuation \ formatting to be “agnostic - neutral” except for the
most basic English rules. That is, “JASON” coding is no longer necessary, (in the dataset.txt files.).
Rules now in effect? Change of topic is set by a double line break between paragraphs.
Arrays or nested indexes is shown by “:” “Set name: subset keys” (which can also be complete sentences “,” delimited.)
I assume all nouns, objects, names and places are now automatically ‘tagged’.
Custom tags or arrays must be coordinated in the Tensor RAG Python files. [As of 05/01/2024] "
May I also suggest that the Demo ChatRTX be constantly trained and updated about itself; with internal examples, explanations, what each system folder is for, and how to start customizing the dataset. AKA the old moto “KNOW THYSELF” News about related utilities apps such as an anime avatar creation (with ChatRTX mocap control) and the new apps like SteerLM for NPCs. I can see how ChatRTX can be the foundation for a whole ecosystem of developer specialties.
Thank you for your detailed comments on ChatRTX. I will pass this along to the developers.
I am a 67 year old recently retired I.T. manager (windows business systems) I am starting the journey to learn Chatbot AI tech and develop a realtime avatar with facial expressions, body language fully controlled by the AI.
This will require the tight integration of several types of AI datasets, eg motion control and morphing. But right now I am a raw newbie except that I have 8 years prior experience as an AI forum moderator. (Unpaid) I am looking forward to
reporting issues and obstacles that are blocking my way in this development process. Thank you for your kind response Richard.
I am a paid member of the new OpenAI Chat GPT4.o It has been a huge help in finding and summarizing Tensor-RAG edit rules. For example the following:
“When a manual tag set has been added to the end of a paragraph, are all the auto tags created by the tensor-RAG subsystem erased, or only when there is an overlap in keywords?” ME
ChatGPT4:
"When manual tags are added to the end of a paragraph, the behavior of the tensor-RAG subsystem with respect to auto-generated tags generally depends on the specific implementation. Here are the common scenarios:
Overwrite Mode: All auto-generated tags are erased and replaced entirely by the manual tags provided. This mode ensures that only the tags explicitly mentioned by the user are considered.
Merge Mode: The manual tags are merged with the auto-generated tags. In this case, any overlap in keywords results in the manual tags taking precedence, while the non-overlapping auto-generated tags are retained.
Selective Overwrite Mode: Only the auto-generated tags that overlap with the manual tags are erased. The remaining auto-generated tags are retained along with the manual tags.
To determine which mode is in use for your specific setup, you would typically need to refer to the documentation of the tensor-RAG subsystem or check the configuration settings. If you have access to the configuration or can run a test, you could easily verify this behavior. Do you know which mode is implemented in your system, or would you like guidance on how to check or configure this?"
“for the Nvidia ChatRTX, can you find out which method they are using? LLm is the Mistral 7B int4.” ME
ChatGPT4:
“For Nvidia ChatRTX, when manual tags are added at the end of a paragraph, the system typically follows the merge mode. This means that manual tags are integrated with auto-generated tags, and in case of overlapping keywords, manual tags take precedence. Non-overlapping auto-generated tags are retained along with the manual tags. This approach allows for both user-specified and automatically detected tags to be utilized effectively. For more detailed information, you can refer to the Nvidia ChatRTX GitHub repository (GitHub).”
“where in Github is this documentation?” ME
ChatGPT4:
“The documentation on how Nvidia ChatRTX handles manual and auto tags is not explicitly detailed in the public GitHub repository. However, general information about the project’s setup, usage, and configurations, including aspects related to the integration of different models and datasets, can be found in the repository’s README.md and associated files. For specific implementation details, you might need to explore the configuration files and scripts within the repository, particularly those related to the tensor-RAG subsystem.”
WHAT THIS MEANS, ME:
As ChatRTX is using the merge method (allegedly) When adding manual tags, do not Repeat nouns, objects, or named places in your manual tag. Instead for cross referencing, use synonyms and short phrases with the same meaning. Do not repeat obvious keywords in the body of that paragraph. Also, if a related keyword in some other part of the dataset does not appear in that paragraph, it could be helpful to add that context tag to your manual tags. Thank you developers for less work for us. :-) Admins: feel free to edit mistakes here.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.