Skip to content

Commit f4f3ee0

Browse files
fixing style, quality, etc in the code
1 parent 958a60f commit f4f3ee0

File tree

17 files changed

+137
-830
lines changed

17 files changed

+137
-830
lines changed

docs/source/_toctree.yml

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -90,6 +90,8 @@
9090
title: LoKr
9191
- local: package_reference/lora
9292
title: LoRA
93+
- local: package_reference/osf
94+
title: OSF
9395
- local: package_reference/xlora
9496
title: X-LoRA
9597
- local: package_reference/adapter_utils

docs/source/package_reference/osf.md

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -105,6 +105,10 @@ config = OSFConfig(
105105
"gate_proj": 4 # Lower rank for gate projection
106106
}
107107
)
108+
109+
# Fractional preserved rank is supported (interpreted per-target as fraction * min_dim)
110+
config = OSFConfig(effective_rank=0.8) # preserve 80% of min_dim; train remaining 20%
111+
config = OSFConfig(rank_pattern={"q_proj": 0.5}) # preserve 50% on q_proj, others use global/default
108112
```
109113

110114
Note: OSF's `effective_rank` is the preserved (frozen) rank, not the trainable rank. The trainable rank equals `min(weight.shape) - effective_rank`. This differs from LoRA's `r`, which directly specifies the trainable rank.

method_comparison/MetaMathQA/experiments/osf/llama-3.2-3B-default/adapter_config.json

Lines changed: 0 additions & 20 deletions
This file was deleted.

method_comparison/MetaMathQA/experiments/osf/llama-3.2-3B-default/training_params.json

Lines changed: 0 additions & 6 deletions
This file was deleted.

method_comparison/MetaMathQA/results/osf--llama-3.2-3B-default.json

Lines changed: 0 additions & 349 deletions
This file was deleted.

0 commit comments

Comments
 (0)