bec864038b
feat: increase max tokens and new tokens in evaluation scripts
6 months ago
424459d840
feat: update evaluation scripts to enhance model configuration and dataset loading, including increased max tokens and added logging
6 months ago
504f0c6c8e
feat: update reward_em_chunk to match only the LAST required paragraph of the reasoning chain and adjust related tests
6 months ago
358875a035
feat: enhance reward_em_chunk function to match multiple paragraphs, add test
6 months ago
2df9f39fda
feat: update model configuration (longer context) and dataset loading logic for improved performance and flexibility
6 months ago
4a1d45271d
feat: add scripts for musique data processing
6 months ago
eebf914a81
refactor: moved modules from src/deepsearch to src/
6 months ago
0f662d4330
refactor: moved FlashRAG submodule from src/ to third_party/
6 months ago
55f34b8503
feat: add FlashRAG as submodule
6 months ago
2fec4f2f42
refactor: change repo stucture (move code from src/ to src/deepsearch)
6 months ago
010957cd99
feat: disable randomization option to get_qa_dataset function by default
6 months ago
1a18cd7bfd
feat: update training and evaluation configurations (editable agent generation scripts)
...
Increased max_generations parameter in agentic_generate and run_eval functions for improved output flexibility.
6 months ago
c8714e0f6b
feat: enhance reward_retry function to handle missing answer tags
...
Added logic to return 0 if the final message from the assistant does not contain answer tags (no matter how hard you try, you won't get anything if no result 💀 )
6 months ago
4de31e0f30
feat: expand reward functions with new strategies and diversity checks
...
- Added reward functions for search strategy and search diversity
- Updated reward_format to include validation for proper message endings.
6 months ago
d0e6068055
fix: strengthen reward correctness logic to handle final message is not asnwer form assistant. Also update logs for reward functions for better debug
...
- Added 'logs/' directory to .gitignore to exclude log files.
- Introduced log_chat_state function to log chat states and rewards to JSONL files.
- Updated reward functions to log chat states with validation results for better tracking and debugging.
6 months ago
338655e563
feat: refine user prompt logic for improved clarity and structure
6 months ago
6d994feeb2
feat: enhance evaluation scripts for base and LoRA models
6 months ago
af7f38c792
feat: add code for qwen architecture
6 months ago
9009440663
chore: disable logging, enable torch complie
6 months ago
d2f03b96ab
feat: enhance evaluation script and remove deprecated shell script
...
- Updated eval.py to streamline model evaluation using vLLM and unsloth.
- Deleted eval.sh as its functionality is now integrated into eval.py.
- Updated .gitignore to exclude eval_logs directory.
6 months ago
31dcbf5d8a
feat: refactor whole code base, add logic for training R1 distil base models, change some template and reward logics
...
- Break down rl_helpers into smaller modules
- Removed deprecated rl_helpers module to streamline the codebase.
- Enhance initial user prompt template inspired by Search-R1
6 months ago
c90c03267e
feat: change user prompt template to search-r1 inspried format
...
use <search></search> instead of embed whole tool definition, which resulted in lots or parsing errors
6 months ago
04593fa8fd
style: change line length to 119, organize imports
6 months ago
3c2deaced9
refactor: restructure code base, better centralize logging logic
7 months ago
thinhlpg