| """Model-independent semantic prompts and state contracts. No labels are accepted.""" |
| import hashlib,json,re |
| from pathlib import Path |
| PROTOCOL=json.loads(Path(__file__).with_name('protocol.json').read_text()) |
| SYSTEM='''You control a camera to search for the object specified by the instruction. Center the requested object in the current view, then submit the current camera pose. |
| Actions: |
| <answer>rotate(dyaw,dpitch)</answer>: rotate relative to the CURRENT pose, using integer degrees. Positive yaw turns right; positive pitch looks up. dyaw must be in [-180,180], dpitch in [-90,90]. |
| <answer>submit(yaw,pitch)</answer>: finish the search. These integers must equal the CURRENT camera pose, not a guessed target location. Submit only when the requested object is centered. |
| Example: at pose (0,0), rotate(-45,0) changes the pose to (315,0). If the target is then centered, the next action is submit(315,0). |
| You may explain your decision before the answer. Return exactly ONE action in exactly ONE <answer>...</answer> block. A maximum of 10 decisions, including submit, is allowed. Use only the supplied observations. Do not invent observations or future events.''' |
| TRACK_TEXT={ |
| 'vision_only':'Available observations: camera images only. No audio is supplied.', |
| 'vision_mono':'Available observations: camera images and the complete 10-second mono recording. The recording is fixed in time and microphone orientation. Mono audio has no left/right channel cues.', |
| 'vision_stereo':'Available observations: camera images and the complete 10-second stereo recording. Left and right channels preserve their ordering. The recording is fixed in time and microphone orientation; camera rotation does not rotate the recording.'} |
| def digest(value):return hashlib.sha256(json.dumps(value,sort_keys=True,ensure_ascii=False,separators=(',',':')).encode()).hexdigest() |
| def messages(instruction,pose,history_actions,decision,track): |
| assert track in TRACK_TEXT and 1<=decision<=10 and len(history_actions)<=2 |
| assert len(pose)==2 and all(isinstance(x,int) for x in pose) |
| content=[] |
| for action in history_actions: |
| content += [dict(type='text',text='Historical observation:'),dict(type='image'),dict(type='text',text='Executed action: '+action)] |
| content += [dict(type='text',text='Current observation:'),dict(type='image')] |
| if track!='vision_only':content.append(dict(type='audio',source='canonical_10s_recording')) |
| content.append(dict(type='text',text=f'Instruction: {instruction}\nCurrent camera pose: ({pose[0]},{pose[1]}).\nDecision: {decision}/10.\n'+TRACK_TEXT[track])) |
| return [dict(role='system',content=SYSTEM),dict(role='user',content=content)] |
| def parse_action(raw,pose): |
| blocks=re.findall(r'<answer>(.*?)</answer>',raw,re.S) |
| if len(blocks)!=1:return None |
| m=re.fullmatch(r'\s*(rotate|submit)\((-?\d+),\s*(-?\d+)\)\s*',blocks[0]) |
| if not m:return None |
| op,y,p=m[1],int(m[2]),int(m[3]) |
| if op=='rotate' and not(-180<=y<=180 and -90<=p<=90):return None |
| if op=='submit' and [y,p]!=pose:return None |
| return op,y,p |
|
|
| def verify_same_state(records): |
| """Compare records for the SAME episode/state/history and same modality track.""" |
| assert records |
| for k in ['episode_id','state_hash','track','protocol_sha256','semantic_prompt_sha256','rgb_sha256','audio_source_sha256','audio_effective_sha256']: |
| assert len({json.dumps(r[k],sort_keys=True) for r in records})==1, 'input/prompt mismatch: '+k |
| return True |
| if __name__=='__main__': |
| a=messages('Find the bell.',[0,0],[],1,'vision_stereo') |
| assert a==messages('Find the bell.',[0,0],[],1,'vision_stereo') |
| assert parse_action('<think>done</think><answer>submit(0,0)</answer>',[0,0])==('submit',0,0) |
| assert parse_action('<answer>submit(1,0)</answer>',[0,0]) is None |
| assert parse_action('<answer>rotate(1,0)</answer><answer>submit(0,0)</answer>',[0,0]) is None |
| print('Common semantic prompt and action-contract checks passed') |
|
|
