File size: 4,290 Bytes
c6f3061 | 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 | """Bounded, inspectable span rewriting with exact preservation of surrounding text."""
import math
import time
import torch
from torch.nn import functional as F
from .model import BOS,EOS,MASK
from .inference import denoise
from .train import amp
@torch.inference_mode()
def local_nll(model,tokenizer,text,start,end,precision="fp32"):
encoded=tokenizer.encode(text)
if len(encoded.ids)+1>model.config.max_length:
raise ValueError("Text exceeds model context")
positions=[i for i,(a,b) in enumerate(encoded.offsets) if b>start and a<end]
if not positions:raise ValueError("Selection has no tokens")
last=min(len(encoded.ids),positions[-1]+1+16)
device=next(model.parameters()).device
tokens=torch.tensor([[BOS]+encoded.ids],device=device)
targets=torch.tensor(encoded.ids+[EOS],device=device)
selected=torch.zeros_like(tokens,dtype=torch.bool)
selected[0,positions[0]:last]=True
with amp(device.type,precision):
logits=model(tokens,torch.zeros(1,device=device),selected,causal=True,right_padded=True)
return float(F.cross_entropy(logits.float(),targets[selected[0]]))
@torch.inference_mode()
def rewrite_span(model,tokenizer,text,start,end,seed=42,precision="fp32"):
if not isinstance(text,str) or not 1<=len(text)<=3000:
raise ValueError("Use a short English text (1..3000 characters)")
if not isinstance(start,int) or not isinstance(end,int) or not 0<=start<end<=len(text):
raise ValueError("Select a passage inside the text")
if not isinstance(seed,int) or not 0<=seed<2**63-16:
raise ValueError("Invalid seed")
original=text[start:end]
a=start+len(original)-len(original.lstrip())
b=end-(len(original)-len(original.rstrip()))
if a>=b or len(text[a:b].split())>12:
raise ValueError("Select 1 to 12 words")
if (a>0 and text[a-1].isalnum() and text[a].isalnum()) or (b<len(text) and text[b-1].isalnum() and text[b].isalnum()):
raise ValueError("Select complete words")
if len(tokenizer.encode(text).ids)>model.config.max_length-32:
raise ValueError("Text too long; shorten it before editing")
prefix,suffix=text[:a],text[b:]
left=tokenizer.encode(prefix.rstrip()).ids
right=tokenizer.encode(suffix).ids
count=len(tokenizer.encode((' ' if prefix else '')+text[a:b]).ids)
if not 1<=count<=16:raise ValueError("Select a shorter passage (at most 16 subword tokens)")
before=local_nll(model,tokenizer,text,a,b,precision)
candidates=[];seen={text[a:b]};started=time.perf_counter()
# Six deterministic attempts; all unique candidates are retained with disclosed ranking.
lengths=[count,count,count,max(1,count-1),min(16,count+1),count]
for attempt,length in enumerate(lengths):
ids=[BOS]+left+[MASK]*length+right+[EOS]
result=denoise(model,tokenizer,ids,steps=12,temperature=0 if attempt==0 else .8,
seed=seed+attempt,precision=precision,allow_eos=False)
fragment=tokenizer.decode(result["tokens"][1+len(left):1+len(left)+length]).strip()
if not fragment or fragment in seen or '\ufffd' in fragment or len(fragment.split())>14:
continue
seen.add(fragment)
rewritten=prefix+fragment+suffix
score=local_nll(model,tokenizer,rewritten,a,a+len(fragment),precision)
trace=[]
for row in result["trace"]:
# The raw model trace is shown for research inspection, not used to rewrite context.
trace.append({"step":row["step"],"remaining":row["remaining"],"model_text":row["text"]})
candidates.append({"replacement":fragment,"text":rewritten,"local_nll":score,
"delta_nll":score-before,"seed":seed+attempt,"masked_tokens":length,"trace":trace})
candidates.sort(key=lambda c:c["local_nll"])
return {"original":text,"selection":{"start":a,"end":b,"text":text[a:b]},
"prefix":prefix,"suffix":suffix,"original_local_nll":before,"candidates":candidates,
"attempts":len(lengths),"latency_ms":(time.perf_counter()-started)*1000,
"method":"Bidirectional denoising; rank by causal NLL of the span and up to 16 following tokens.",
"limitation":"Ranking reflects this model's preferences, not verified grammar or semantic correctness."}
|