CBIT GPT

更新日志

Upgradelog

Add: Logger integrated into model fine-tuning process.
Add: Support for fine-tuning multiple models simultaneously and comparing results
Upgrades: Optimized reward model mechanism for online outputs, incorporating the latest prompts
Fixes:Optimized the maximum length of conversations

New: Implemented a Hybrid Training Pipeline that integrates OpenAI’s GPT-4o, enabling the generation of multimodal domain outputs. *This approach minimizes the need for computational resources and enhances the efficiency of output delivery time.
Add: Optimized temporal dimension tags by incorporating time-embedded vectors within the model’s architecture, and added time-stamped indexes to the vector database.
Add: Implemented an export feature, allowing the export of chat history to JSON format.
Fix: Resolved issues encountered when integrating models with RESTful API.

New: Comprehensive upgrade and fine-tuning based on the Llama3 model with hybrid training techniques.
Added: Enhanced capabilities for file search and vector storage solutions.
Developed and deployed RESTful API infrastructure.

Strengthened security mechanisms for user vector databases, including key pairing enhancements.
Fix: Resolved concurrency conflicts in search engine invocation.
Fix: Streamlined the embeddings system upgrade process for improved efficiency.

Fix: Addressed improper handling of banned words in certain scenarios.
Fix: Corrected the issue where Embeddings Auto-Sync was not properly initiating the finetune model.
Fix: Remedied the malfunctions in the Audio to Text conversion process.

Update: Advanced the Search Engine embeddings environment to support contextual analysis and made updates to the models and vectors table.
Update: Refactored the Simple API for enhanced consistency across endpoints.
Fix: Addressed a security vulnerability related to arbitrary file uploads.

Fix: Addressed several PHP notices and warnings to enhance code stability.
Add: Implemented a routine to automatically purge outdated embeddings from the database for optimal performance.