File size: 7,599 Bytes
d35dd87 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 | """Inference-only development diagnostics; original labels are reference labels.
Residue shuffling is an out-of-distribution intervention, not a label-preserving
biological transformation. Original calibration temperatures remain fixed.
"""
from __future__ import annotations
import argparse
from collections import Counter
import gc
import hashlib
import json
from pathlib import Path
import random
import numpy as np
import torch
import laya_formal_experiment as core
from laya_direct_bpe import DirectBPE, digest
from laya_direct_experiment import make_items
ROOT=core.ROOT
SEEDS=(20260922,20260923,20260924)
def perturbed(row,variant):
row=dict(row)
if variant=='sequence_removed':row['sequence']=''
elif variant.startswith('residue_shuffle_'):
raw=row['sequence'];characters=list(raw)
seed=int.from_bytes(hashlib.sha256(f"diagnostic-v1:{variant}:{row['id']}".encode()).digest()[:8],'big')
random.Random(seed).shuffle(characters)
row['sequence']=''.join(characters)
if Counter(raw)!=Counter(row['sequence']):raise ValueError('Shuffle changed residue composition')
return row
def permutation(row):
order=list(range(len(row['choices'])))
seed=int.from_bytes(hashlib.sha256(f"candidate-diagnostic-v1:{row['id']}".encode()).digest()[:8],'big')
random.Random(seed).shuffle(order)
if order==list(range(len(order))):order=order[1:]+order[:1]
return order
def canonicalize(records,orders):
result=[]
for record in records:
order=orders[record['id']]
logits=[0.]*len(order)
for position,canonical in enumerate(order):logits[canonical]=record['logits'][position]
item=dict(record,logits=logits,label=order[record['label']])
values=np.asarray(logits,dtype=float);exponent=np.exp(values-values.max())
item['probs']=(exponent/exponent.sum()).tolist()
result.append(item)
return result
def compare(reference,changed):
if [(r['id'],r['label'],r['task']) for r in reference]!=[(r['id'],r['label'],r['task']) for r in changed]:
raise ValueError('Diagnostic membership/labels changed')
out={}
for task in sorted({r['task'] for r in reference}):
pairs=[(a,b) for a,b in zip(reference,changed) if a['task']==task]
p=np.asarray([a['probs'] for a,b in pairs]);q=np.asarray([b['probs'] for a,b in pairs])
if not np.isfinite(p).all() or not np.isfinite(q).all():raise FloatingPointError('Non-finite diagnostic')
middle=(p+q)/2
js=.5*np.sum(p*np.log(np.maximum(p,1e-300)/np.maximum(middle,1e-300))+
q*np.log(np.maximum(q,1e-300)/np.maximum(middle,1e-300)),axis=1)
out[task]={'n':len(pairs),'prediction_agreement':float((p.argmax(1)==q.argmax(1)).mean()),
'mean_probability_l1':float(np.abs(p-q).sum(1).mean()),
'mean_max_probability_difference':float(np.abs(p-q).max(1).mean()),
'mean_js_divergence':float(js.mean())}
return out
def run_one(run_dir,out,batch_size=16):
if (out/'summary.json').exists():return json.loads((out/'summary.json').read_text())
saved=json.loads((run_dir/'summary.json').read_text())
if not saved['formal'] or saved['test_access'] or not saved['checkpoint_reload_logits_match']:
raise ValueError('Source checkpoint failed protocol gate')
rep=DirectBPE(run_dir/'checkpoint/representation')
eligible=core.load_eligible_ids(saved['eligible_id_filter'])
rows=[r for r in core.load_split(ROOT/'artifacts/laya_formal_data','both','selection_dev') if r['id'] in eligible]
_,build_model,builder=core.import_laya('vendor/laya')
cfg=json.loads((run_dir/'checkpoint/rl_agent_config.json').read_text())
model=core.fresh_reload(run_dir/'checkpoint',cfg,build_model,torch.device('cuda'))
tokenizer=rep.expanded if saved['condition']=='full_bpe' else rep.base
base_items=make_items(rows,rep,saved['condition'],builder,saved['seed'])
baseline=core.evaluate(model,base_items,tokenizer,torch.device('cuda'),batch_size)
previous=[json.loads(x) for x in (run_dir/'selection_dev_predictions.jsonl').read_text().splitlines()]
if len(previous)!=len(baseline) or any(a['id']!=b['id'] or max(abs(x-y) for x,y in zip(a['logits'],b['logits']))>1e-5
for a,b in zip(previous,baseline)):
raise ValueError('Unperturbed predictions do not reproduce source result')
temps={t:v['temperature'] for t,v in saved['calibration_temperature'].items()}
result={'condition':saved['condition'],'seed':saved['seed'],'split':'selection_dev','test_access':False,
'no_training':True,'reference_labels_not_assumed_valid_under_sequence_intervention':True,
'source_summary_sha256':digest(run_dir/'summary.json'),
'baseline_logits_reproduced':True,'temperature_refitted':False,
'original':{'raw':core.metric_from_records(baseline),'calibrated':core.metric_from_records(baseline,temps)},
'variants':{}}
out.mkdir(parents=True,exist_ok=True)
for variant in ('sequence_removed','residue_shuffle_0','residue_shuffle_1','residue_shuffle_2','candidate_permutation'):
if variant=='candidate_permutation':
items=[];orders={}
for row in rows:
order=permutation(row);orders[row['id']]=order
ids,markers=rep.build(row,saved['condition'],builder,order)
items.append({'ids':ids,'markers':markers,'label':order.index(row['label']),
'task':row['task'],'id':row['id']})
predictions=canonicalize(core.evaluate(model,items,tokenizer,torch.device('cuda'),batch_size),orders)
unchanged=None
else:
changed=[perturbed(row,variant) for row in rows]
unchanged=sum(a['sequence']==b['sequence'] for a,b in zip(rows,changed))
items=make_items(changed,rep,saved['condition'],builder,saved['seed'])
predictions=core.evaluate(model,items,tokenizer,torch.device('cuda'),batch_size)
result['variants'][variant]={'reference_label_metrics':{'raw':core.metric_from_records(predictions),
'calibrated':core.metric_from_records(predictions,temps)},
'comparison_to_original':compare(baseline,predictions),
'unchanged_sequence_count':unchanged}
(out/f'{variant}_predictions.jsonl').write_text(''.join(json.dumps(r)+'\n' for r in predictions))
print(json.dumps({'condition':saved['condition'],'seed':saved['seed'],'variant':variant,
'comparison':result['variants'][variant]['comparison_to_original']}),flush=True)
(out/'summary.json').write_text(json.dumps(result,indent=2)+'\n')
del model
gc.collect();torch.cuda.empty_cache()
return result
def main():
p=argparse.ArgumentParser(description=__doc__)
p.add_argument('--output-dir',type=Path,default=ROOT/'artifacts/laya_controls/sequence_diagnostics')
p.add_argument('--seeds',type=int,nargs='+',default=list(SEEDS))
p.add_argument('--conditions',nargs='+',choices=['raw','full_bpe'],default=['full_bpe','raw'])
args=p.parse_args()
for seed in args.seeds:
for condition in args.conditions:
run_one(ROOT/f'artifacts/laya_direct_legacy/{condition}_seed{seed}',
args.output_dir/f'{condition}_seed{seed}')
print('All requested development diagnostics completed.',flush=True)
if __name__=='__main__':main()
|