Skip to content

Distillation v1#18

Open
bl4ckse4m wants to merge 13 commits into
mainfrom
distillation
Open

Distillation v1#18
bl4ckse4m wants to merge 13 commits into
mainfrom
distillation

Conversation

@bl4ckse4m
Copy link
Copy Markdown
Collaborator

Initial working version of distillation pipeline with severe quality issues

# Conflicts:
#	inference_and_eval/musique_inference_with_db.py
#	setup_db.sh
#	src/wikontic/utils/openai_utils.py
…around for trl reading wrong encoding on windows machine
# Conflicts:
#	inference_and_eval/hotpot_inference_with_db.py
#	inference_and_eval/mine_inference.py
bl4ckse4m added 7 commits May 17, 2026 23:43
Small analysis notebook.
Training now has a curriculum scheduler on triplets and loss down-weighting for "instance of".
Overall full-match F1 0.55
…, trained longer for 5 epochs

Added cosine scheduler and enabled relation weighting (besides instance_of)
As a result full-match F1 pushed from 0.55 to 0.656
…y overlapping;

Picked a separate test set from data;
Full match F1 went back down to 0.557
Added how_to_run instruction
…ts with focal loss

Increased lora_r to 32 from 16, added gate_proj, up_proj, down_proj to LoRA targets
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants