|
Download trace/notebook.md from Perfect7613/weightpatch-run-math-gsm8k: direct link, hf CLI and curl.
- Browser
- Download file 797 Bytes
-
https://huggingface.co/datasets/Perfect7613/weightpatch-run-math-gsm8k/resolve/main/trace/notebook.md
- Command line
-
hf download hf://datasets/Perfect7613/weightpatch-run-math-gsm8k/trace/notebook.md
-
curl -L -o notebook.md https://huggingface.co/datasets/Perfect7613/weightpatch-run-math-gsm8k/resolve/main/trace/notebook.md
797 Bytes
| - Read atlas: parallel attention/MLP, neuron writes are down_proj columns; base reference is not a better checkpoint. Reserve math items 128:256 for validation. No edits yet. | |
| - Finished after two unsuccessful hypothesis rounds; all edits reverted. Round1 L33N10488 LaTeX writer: half fixed0/broke1; zero fixed1/broke1, paired CI [-0.0234,0.0234]. Round2 L34N5892 planning writer 0.75: fixed1/broke2, CI [-0.0312,0.0156]. Discovery baseline101/128. No significant improvement, so no submission. Final full math: {'n': 256, 'accuracy': 0.765625, 'released_accuracy': 0.765625, 'vs_released': {'diff': 0.0, 'lo': 0.0, 'hi': 0.0, 'n': 256, 'significant': False, 'fixed': 0, 'broke': 0}, 'unfinished': 24, 'looping': 0, 'released_unfinished': 24} Final safety: harmful=0.75, benign=0.020833333333333332 |