randomfolk's picture
Add anonymous scratch checkpoints, synthetic data, and source
86316a2 verified
Raw History Blame Contribute Delete
5.6 kB
"""Uncached batched causal choice scoring, validation maps only."""
import argparse
import hashlib
import json
import re
import time
from pathlib import Path
import numpy as np
from mechcal.training.spatial_transcripts import render_map, LETTERS
def main():
ap=argparse.ArgumentParser()
ap.add_argument('--weight',required=True)
ap.add_argument('--mode',choices=['full','choice_only'],required=True)
ap.add_argument('--smoke',action='store_true')
ap.add_argument('--batch-size',type=int,default=4)
ap.add_argument('--test-main',action='store_true')
args=ap.parse_args()
import unsloth
import torch
from unsloth import FastLanguageModel
from trl import DataCollatorForCompletionOnlyLM
name='reward_w'+args.weight
source=Path('data/spatial_llama_eval_v1')/(name+'.json')
if args.test_main: source=Path('data/llama_test_20260925/task_b')/(name+'.json')
data=json.loads(source.read_text()); records=data['records']; maps=np.array(data['donor_maps'])
adapter=Path('outputs/spatial_llama_v1')/(name+'_'+args.mode+'_seed100')
out=Path('outputs/spatial_llama_eval_fixedshape_v2')/('smoke' if args.smoke else 'production')/name/args.mode
if args.test_main: out=Path('outputs/llama_test_20260925/task_b')/('smoke' if args.smoke else 'production')/name/args.mode
out.mkdir(parents=True,exist_ok=True)
model,tok=FastLanguageModel.from_pretrained(model_name=str(adapter),max_seq_length=32768,dtype=None,load_in_4bit=True)
FastLanguageModel.for_inference(model); model.eval()
tok.pad_token_id=0; tok.padding_side='right'
collator=DataCollatorForCompletionOnlyLM(response_template=tok(' <<').input_ids[1:],instruction_template=tok('>>').input_ids[1:],tokenizer=tok)
choices=[tok.encode(c,add_special_tokens=False) for c in LETTERS]
assert all(len(c)==1 for c in choices)
choices=[c[0] for c in choices]; device=next(model.parameters()).device
def encode(text):
enc=tok(text,truncation=False)
b=collator([enc]); pos=b['labels'][0].ne(-100).nonzero().flatten()
acts=np.array([LETTERS.index(c) for c in re.findall(r'<<([A-Y])>>',text)])
assert len(pos)==20 and np.array_equal(b['input_ids'][0,pos].numpy(),np.array(choices)[acts])
return b['input_ids'][0],pos
@torch.inference_mode()
def forward(batch):
# Identical shapes for every intact/donor call, including smoke checks.
# This controls quantized-kernel rounding driven by padding length.
length=1024
assert len(batch)<=args.batch_size and all(len(x[0])<=length for x in batch)
padded=batch+[batch[0]]*(args.batch_size-len(batch))
ids=torch.zeros((args.batch_size,length),dtype=torch.long,device=device)
mask=torch.zeros_like(ids)
for i,(tokens,pos) in enumerate(padded): ids[i,:len(tokens)]=tokens.to(device); mask[i,:len(tokens)]=1
result=model(input_ids=ids,attention_mask=mask,use_cache=False,return_dict=True)
lp=torch.stack([result.logits[i,(pos-1).to(device)].float().log_softmax(-1)[:,choices] for i,(_,pos) in enumerate(batch)]).cpu().numpy()
assert np.isfinite(lp).all() and np.all(np.exp(lp).sum(-1)<=1.0001)
return lp
# Same-shape future-token perturbation must not affect the first prediction.
probe=encode(render_map(records[0]['rounds'][0],args.mode))
changed=probe[0].clone(); changed[probe[1][0]:]=choices[0]
checks=forward([probe,(changed,probe[1])])
error=float(np.max(np.abs(checks[0,0]-checks[1,0])))
assert error < .003, ('causal prefix check',error)
start=time.time()
for i in range(2 if args.smoke else 50):
target=out/f'participant_{i:03d}.npz'
if target.exists(): continue
episodes=records[i]['rounds']
texts=[render_map(ep,args.mode) for ep in episodes]
repeats=2 if args.smoke else 20
if args.mode=='full':
for r in range(repeats):
for j in (6,7):
ep=dict(episodes[j]); ep['rewards']=records[int(maps[r,i])]['rounds'][j]['rewards']
texts.append(render_map(ep,args.mode))
encoded=[encode(t) for t in texts]
values=np.concatenate([forward(encoded[k:k+args.batch_size]) for k in range(0,len(encoded),args.batch_size)])
payload=dict(index=i,participant=str(records[i]['participant']),actions=np.array(data['actions'][i]),donor_indices=maps[:repeats,i],intact_logp=values[:8])
if args.mode=='full':
payload['donor_logp']=values[8:].reshape(repeats,2,20,25)
first_error=float(np.max(np.abs(payload['donor_logp'][:,:,0]-values[None,6:8,0])))
assert first_error < .003, ('actual donor unchanged-prefix check',i,first_error)
payload['first_choice_max_logp_difference']=first_error
temporary=target.with_suffix('.partial.npz')
np.savez_compressed(temporary,**payload)
temporary.replace(target)
print(f'COMPLETE participant={i} elapsed={time.time()-start:.1f}',flush=True)
manifest=dict(complete=True,smoke=args.smoke,test_used=args.test_main,split='test' if args.test_main else 'val',weight=args.weight,mode=args.mode,causal_max_difference=error,batch_size=args.batch_size,fixed_length=1024,source_sha256=hashlib.sha256(source.read_bytes()).hexdigest(),script_sha256=hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),adapter_config_sha256=hashlib.sha256((adapter/'adapter_config.json').read_bytes()).hexdigest())
(out/'COMPLETE.json').write_text(json.dumps(manifest,indent=2))
if __name__=='__main__': main()