MIT LICENSE, README license fix, make_dataset.py, 019 re-evaluation, technical report and CITATION.cff
86415de verified | import os,subprocess,sys,glob,shutil | |
| subprocess.run([sys.executable,'-m','pip','install','-q','transformers==4.46.3'],check=True) | |
| inp=os.path.dirname(glob.glob('/kaggle/input/**/k2long_s17.pt',recursive=True)[0]) | |
| S='/kaggle/working/src'; os.makedirs(S+'/data',exist_ok=True); shutil.copy(f'{inp}/expanded.json',f'{S}/data/') | |
| shutil.copy(f'{inp}/run018.py','/kaggle/working/'); open('/kaggle/working/diag_pad.py','w').write("# Diagnostic: does the text-baseline score depend on batching/padding or attention kernel?\nimport os,sys,json,torch\nsys.path.insert(0,os.path.dirname(os.path.abspath(__file__)))\nimport run018 as R\nSUB=R.FRESH['id'][:48]+R.FRESH['ood'][:48]\nout={'torch':torch.__version__,'device':R.DEV,'attn':R.lm.config._attn_implementation}\nfor bs in (1,24):\n out[f'past_bs{bs}']=R.acc('text',SUB,R.SUF_HELD['past'],bs=bs)\n out[f'phrase_bs{bs}']=R.acc('text',R.TRANSFER['1_phrase'],bs=bs)\n out[f'ood_bs{bs}']=R.acc('text',R.FRESH['ood'],bs=bs)\nprint(json.dumps(out),flush=True)\n") | |
| from transformers import AutoTokenizer,AutoModelForCausalLM | |
| mp='/kaggle/working/m135'; R='12fd25f77366fa6b3b4b768ec3050bf629380bac' | |
| AutoTokenizer.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp); AutoModelForCausalLM.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp) | |
| subprocess.run([sys.executable,'/kaggle/working/diag_pad.py'],env=dict(os.environ,SRC=S,MP=mp,OD='/kaggle/working',INP=inp),check=True) | |
| shutil.rmtree(mp); shutil.rmtree(S); os.remove('/kaggle/working/run018.py') | |