mirror/khoj
1
0
Fork 0
mirror of https://github.com/khoj-ai/khoj.git synced 2026-05-19 06:07:32 -04:00
Code Issues Packages Projects Releases Wiki Activity
4,389 Commits 18 Branches 175 Tags
create-data-tracer
Commit Graph

13 Commits

Author SHA1 Message Date
sabaimran
f3045ab436 replace system prompt for full thoughts, not for datatrace 2025-02-03 23:23:33 -08:00
sabaimran
899e9a6c23 Add additional log lines 2025-02-03 13:16:07 -08:00
sabaimran
a35002e93b Use log level info for constructing the dataset. Temp: comment out error lines when reading in JSONL content 2025-02-03 13:04:19 -08:00
sabaimran
3db3d7bbb4 add print line about full thoughts 2025-02-03 07:17:42 -08:00
sabaimran
f870a2f7a2 Add more outputs about nature of dataset when uploading and configure custom system prompt 2025-02-03 07:11:34 -08:00
sabaimran
9815bbc8a1 Add method for loading the entire dataset, loading good evaluated rows, and uploading a merged dataset 2025-01-30 22:39:46 -08:00
sabaimran
06a900069f add line by line parsing logic for the dataset jsonl file 2025-01-29 11:58:32 -08:00
sabaimran
79f3f61b70 allow multiple eval_paths to combine eval results across multiple runs 2025-01-27 15:29:29 -08:00
sabaimran
6257d1bb62 Extract good rows before attempting to upload full thoughts df 2025-01-26 22:33:07 -08:00
sabaimran
ef794a78d5 Publish thoughts dataset to a separate repo from the subiteration traces 2025-01-26 21:50:03 -08:00
sabaimran
e5d5153fc4 Allow setting dynamic filepaths for full thought trace. Update data upload script to handle full thoughts. 2025-01-26 21:44:26 -08:00
sabaimran
7f001768c1 Update output and input formats for datasets 
- intermediate trains of thought are outputted as jsonl in order to store nested / complex data
- evaluation results are stored as csv without references for simplicity
- the create dataset file reads both of them, combines and outputs the results ot huggingface, if possible
 2025-01-26 21:42:06 +07:00
Debanjum
818239c3c7 Allow saving Khoj chat actor query, response traces to HF dataset 
- Data tracer can be enabled by setting DATATRACE_PATH env var
- Save query, response pairs to a HF dataset locally in trainable
  format
- Currently only enabled for research mode apick_next_tool research mode
  chat actor
- Use prompt tracer like code execution hooks
 2025-01-26 03:03:17 +07:00