import json,re,collections,difflib
def norm_cmd(c):
  c=re.sub(r'^\[[^\]]+\]\s*','',c); c=c.replace('\\"','"'); return c.strip()
def tmpl(s): return re.sub(r'\d+','N',s.lower())
stats=collections.Counter()
for sp in ['train','val','test']:
  rows=[json.loads(l) for l in open(f'train/intent_v1/{sp}.jsonl')]
  out=[];seen=set()
  for r in rows:
    c=norm_cmd(r['command'])
    if c!=r['command']: stats['cmd_normalized']+=1
    r['command_original']=r['command']; r['command']=c
    it=r['intent'].strip()
    # restate: intent tokens mostly == command tokens
    ct=set(re.findall(r'[a-z0-9]+',c.lower())); itok=re.findall(r'[a-z0-9]+',it.lower())
    if difflib.SequenceMatcher(None,it.lower(),c.lower()).ratio()>0.85 or (itok and all(t in ct for t in itok)):
      stats['restate']+=1;continue
    k=(r['vendor'],tmpl(c))
    if k in seen: stats['templated_dup']+=1;continue
    seen.add(k); out.append(r)
  with open(f'train/intent_v1.1/{sp}.jsonl','w') as f:
    for r in out: f.write(json.dumps(r)+'\n')
  stats[sp]=len(out)
print(dict(stats))