fix: resolve duplicate -g flag and type=bool argparse bug in sft.py

[vominh1919]

Problem

Two bugs in scripts/sft.py's argument parser:

1. Duplicate -g flag: Both --gradient-accumulation-steps (line 69) and --max-grad-norm (line 73) use -g as their short flag. The second definition shadows the first, so python sft.py -g 4 silently sets max_grad_norm=4.0 instead of gradient_accumulation_steps=4.

2. type=bool doesn't work in argparse: bool("False") is True in Python — any non-empty string is truthy. The only way to disable --push-to-hub via CLI is --push-to-hub "", which is unintuitive.

Fix

1. Change --gradient-accumulation-steps short flag from -g to -G (uppercase, since --max-grad-norm already owns -g)
2. Replace type=bool with action=argparse.BooleanOptionalAction, which provides both --push-to-hub and --no-push-to-hub

Before vs After

Scenario	Before	After
sft.py -g 4	Sets max_grad_norm=4.0 (wrong!)	Sets max_grad_norm=4.0 (correct, -G needed for grad accum)
sft.py -G 4	Error: unrecognized argument	Sets gradient_accumulation_steps=4 ✓
sft.py --no-push-to-hub	Error: unrecognized argument	Sets push_to_hub=False ✓
sft.py --push-to-hub False	Sets push_to_hub=True (wrong!)	N/A — use --no-push-to-hub instead

Tests

Verified syntax with python3 -c "import ast; ast.parse(open('scripts/sft.py').read())"

---

Note

Low Risk
Low risk: only adjusts CLI argument definitions in scripts/sft.py, changing how flags are parsed but not training logic or data handling.

Overview
Fixes the scripts/sft.py CLI so --gradient-accumulation-steps no longer conflicts with --max-grad-norm by switching its short flag from -g to -G.

Updates --push-to-hub to use argparse.BooleanOptionalAction, enabling --push-to-hub / --no-push-to-hub instead of the broken type=bool parsing.

^{Reviewed by Cursor Bugbot for commit d2ae717. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d2ae717. Configure here.}

[cursor]

Parsed push_to_hub value never used in SFTConfig

High Severity

The --push-to-hub argparse fix is incomplete. While the CLI now correctly parses --no-push-to-hub via BooleanOptionalAction, the SFTConfig on line 49 hardcodes push_to_hub=True instead of using args.push_to_hub. This means --no-push-to-hub is silently ignored — the model is always pushed to the hub regardless of the flag.

Additional Locations (1)

• scripts/sft.py#L48-L49

 

^{Reviewed by Cursor Bugbot for commit d2ae717. Configure here.}