| |
| import sys |
| from pathlib import Path |
| _ROOT = str(Path(__file__).resolve().parents[2]) |
| if _ROOT not in sys.path: |
| sys.path.insert(0, _ROOT) |
|
|
| from inspect_ai import Task,task |
| from inspect_ai.dataset import Sample,MemoryDataset |
| from inspect_ai.model import get_model,GenerateConfig,ChatMessageUser |
| from inspect_ai.solver import solver |
| from inspect_ai.scorer import scorer,Score,mean |
| from encoding_tax.common import ROOT,load_config,digest |
| from encoding_tax.prompts import serialize,reasoning_stops,continue_answer |
| from encoding_tax.scoring import probability_score |
| from encoding_tax.data import load_rows |
| from extensions.e6_exploratory_v2 import provider |
| from extensions.e6_exploratory_v2.design import require_extension,cases,case_id,truth,body,score_diagnostic,diagnostic_answer |
|
|
| @solver |
| def extension_solver(): |
| base=load_config() |
| async def solve(state,generate): |
| m=state.metadata;p=base['models'][m['reader']] |
| text,trace=body(m,base);raw=serialize(p,text,'scratchpad') |
| output=await get_model().generate([ChatMessageUser(content=raw)],config=GenerateConfig(max_tokens=m['budget'],stop_seqs=reasoning_stops(p),temperature=0,top_p=1,max_retries=0)) |
| m.update(supplied_trace=trace,scratchpad=output.completion,requests=[output.metadata],reasoning_end=output.metadata['finish_reason']) |
| answer_prompt=continue_answer(p,raw,output.completion)+('[' if m['mode']=='diagnostic' else '') |
| m['answer_prefill']='[' if m['mode']=='diagnostic' else '' |
| state.output=await get_model().generate([ChatMessageUser(content=answer_prompt)],config=GenerateConfig(max_tokens=48,stop_seqs=[']'] if m['mode']=='diagnostic' else ['\n'],temperature=0,top_p=1,max_retries=0)) |
| m['requests'].append(state.output.metadata) |
| state.messages=[ChatMessageUser(content=answer_prompt),state.output.message];state.completed=True |
| return state |
| return solve |
|
|
| @scorer(metrics=[mean()]) |
| def extension_score(): |
| async def score(state,target): |
| m=state.metadata;raw=state.output.completion |
| result=probability_score(raw,m['label']) if m['mode']=='monitor' else score_diagnostic(diagnostic_answer(raw) if state.output.metadata.get('finish_reason')=='stop' else '',truth(m)) |
| result['raw_suffix']=raw |
| if m['mode']=='diagnostic':result['reconstructed_answer']=diagnostic_answer(raw) |
| m['outcome']=result |
| return Score(value=result['probability'] if m['mode']=='monitor' else result['all_correct'],answer=raw,metadata=result) |
| return score |
|
|
| def build(reader): |
| cfg,base,pin=require_extension() |
| rows=load_rows({'datasets':{'synthetic':cfg['dataset']}},'synthetic','pilot') |
| data=[Sample(id=case_id(m),input=m['problem'],target=str(m['label']),metadata=m) for m in cases(rows,reader)] |
| model=get_model('encoding_tax_e6x2/'+reader,study_config=cfg['baseline_config'],audit_dir='results/e6-exploratory-v2/pilot/transport',config=GenerateConfig(max_connections=16,max_retries=0)) |
| return Task(name='E6X2_pilot',dataset=MemoryDataset(data),solver=extension_solver(),scorer=extension_score(),model=model,config=GenerateConfig(max_connections=16,max_retries=0,temperature=0,top_p=1),fail_on_error=False,metadata={'exploratory':True,'run_version':'e6x2','reader':reader,'split':'pilot','extension_freeze':pin,'extension_freeze_sha256':digest(ROOT/'extensions/e6_exploratory_v2/freeze.json'),'answer_cap':48,'dataset':cfg['dataset'],'concurrency':16}) |
|
|
| @task |
| def gemma():return build('gemma') |
| @task |
| def gptoss():return build('gptoss') |
| @task |
| def nemotron():return build('nemotron') |
| @task |
| def qwen27():return build('qwen27') |
| @task |
| def qwen2t():return build('qwen2t') |
|
|
