Title: Fix: Shape Mismatch during Left Padding Adjustment in compute_metrics (Generated by Ana - AI SDE)

[ana-ai-sde]

Description:

Fix for Issue 88

This pull request addresses a RuntimeError caused by a shape mismatch during the left padding adjustment in the compute_metrics function of examples/loreft/compute_metrics.py. The issue arises when the left_padding tensor is empty, leading to an incompatible broadcasting operation with the intervention_locations tensor.

This patch was generated by Ana - AI SDE, an AI-powered software development assistant.

The fix introduces a check for the presence of elements in left_padding. If left_padding is empty, a warning message is printed, and the adjustment is skipped. This ensures the compatibility of tensor shapes and prevents the RuntimeError.

This patch improves the robustness of the compute_metrics function by handling edge cases related to left padding.

[frankaging]

@ana-ai-sde hey Ana, thanks for the PR, have you run any test on this change? thanks!

[d4rk-lucif3r]

Hi @frankaging,

I am the author of Ana - AI SDE.

Yes, we did our validations. The test results weren't added as we are still working on the pull request template.

We tried running the same commands as mentioned in Issue 88.

Command:

python train.py -task gsm8k -model /home/Meta-Llama-3-8B-Instruct-function-calling-json-mode -seed 42 -l all -r 4 -p f7+l7 -e 12 -lr 9e-4 -type NodireftIntervention -gradient_accumulation_steps 4 -batch_size 1 -eval_batch_size 1 --dropout 0.05 --test_split validation --use_normalized_template --greedy_decoding --warmup_ratio 0.00 --weight_decay 0.06

Output Without Fix:

Output With Fix:

The runtime error was fixed, the code proceeded to the next steps, and eventually training finished.

If you have any doubts or questions, feel free to ask.

Thanks,
Arsh Anwar

[aryamanarora]

LGTM!