Spaces:
Sleeping
Sleeping
Isolate HF sessions and stabilize hosted baseline
Browse files- inference.py +21 -2
- server.py +26 -7
inference.py
CHANGED
|
@@ -24,6 +24,7 @@ from __future__ import annotations
|
|
| 24 |
|
| 25 |
import asyncio
|
| 26 |
import os
|
|
|
|
| 27 |
import sys
|
| 28 |
import textwrap
|
| 29 |
import traceback
|
|
@@ -179,6 +180,8 @@ def should_accept_rewrite(draft: str, candidate: str) -> bool:
|
|
| 179 |
if not candidate:
|
| 180 |
return False
|
| 181 |
|
|
|
|
|
|
|
| 182 |
draft_features = extract_features(draft)
|
| 183 |
candidate_features = extract_features(candidate)
|
| 184 |
|
|
@@ -186,6 +189,16 @@ def should_accept_rewrite(draft: str, candidate: str) -> bool:
|
|
| 186 |
return False
|
| 187 |
if _count_questions(candidate) > 1 or candidate_features.interrogative > 0:
|
| 188 |
return False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 189 |
|
| 190 |
# Do not let the rewrite weaken the key stage-driving signals already
|
| 191 |
# present in the deterministic draft.
|
|
@@ -199,8 +212,14 @@ def should_accept_rewrite(draft: str, candidate: str) -> bool:
|
|
| 199 |
return False
|
| 200 |
if draft_features.safety > 0 and candidate_features.safety <= 0:
|
| 201 |
return False
|
| 202 |
-
|
| 203 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 204 |
|
| 205 |
|
| 206 |
# -------------------------- per-task episode ---------------------------------
|
|
|
|
| 24 |
|
| 25 |
import asyncio
|
| 26 |
import os
|
| 27 |
+
import re
|
| 28 |
import sys
|
| 29 |
import textwrap
|
| 30 |
import traceback
|
|
|
|
| 180 |
if not candidate:
|
| 181 |
return False
|
| 182 |
|
| 183 |
+
draft_norm = " ".join(re.sub(r"[^\w\s]", "", draft.lower()).split())
|
| 184 |
+
candidate_norm = " ".join(re.sub(r"[^\w\s]", "", candidate.lower()).split())
|
| 185 |
draft_features = extract_features(draft)
|
| 186 |
candidate_features = extract_features(candidate)
|
| 187 |
|
|
|
|
| 189 |
return False
|
| 190 |
if _count_questions(candidate) > 1 or candidate_features.interrogative > 0:
|
| 191 |
return False
|
| 192 |
+
if len(candidate.split()) > max(24, int(len(draft.split()) * 1.2)):
|
| 193 |
+
return False
|
| 194 |
+
if draft_features.open_question != candidate_features.open_question:
|
| 195 |
+
return False
|
| 196 |
+
if draft_features.advice != candidate_features.advice:
|
| 197 |
+
return False
|
| 198 |
+
if draft_features.safety != candidate_features.safety:
|
| 199 |
+
return False
|
| 200 |
+
if draft_features.validation != candidate_features.validation:
|
| 201 |
+
return False
|
| 202 |
|
| 203 |
# Do not let the rewrite weaken the key stage-driving signals already
|
| 204 |
# present in the deterministic draft.
|
|
|
|
| 212 |
return False
|
| 213 |
if draft_features.safety > 0 and candidate_features.safety <= 0:
|
| 214 |
return False
|
| 215 |
+
if draft_norm == candidate_norm:
|
| 216 |
+
return True
|
| 217 |
+
|
| 218 |
+
# Only accept near-verbatim rewrites; otherwise keep the proven draft.
|
| 219 |
+
draft_tokens = set(draft_norm.split())
|
| 220 |
+
candidate_tokens = set(candidate_norm.split())
|
| 221 |
+
overlap = len(draft_tokens & candidate_tokens) / max(1, len(draft_tokens))
|
| 222 |
+
return overlap >= 0.8
|
| 223 |
|
| 224 |
|
| 225 |
# -------------------------- per-task episode ---------------------------------
|
server.py
CHANGED
|
@@ -14,7 +14,10 @@ instances.
|
|
| 14 |
"""
|
| 15 |
from __future__ import annotations
|
| 16 |
|
|
|
|
|
|
|
| 17 |
from fastapi import FastAPI, HTTPException
|
|
|
|
| 18 |
|
| 19 |
from src.env import ESCEnv
|
| 20 |
from src.models import ResetRequest, StepRequest
|
|
@@ -29,7 +32,15 @@ app = FastAPI(
|
|
| 29 |
),
|
| 30 |
)
|
| 31 |
|
| 32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
|
| 35 |
@app.get("/")
|
|
@@ -48,27 +59,35 @@ def list_tasks() -> dict:
|
|
| 48 |
|
| 49 |
|
| 50 |
@app.post("/reset")
|
| 51 |
-
def reset(req: ResetRequest | None = None) -> dict:
|
| 52 |
req = req or ResetRequest()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
try:
|
| 54 |
-
result =
|
| 55 |
except KeyError as e:
|
| 56 |
raise HTTPException(status_code=400, detail=str(e))
|
|
|
|
| 57 |
return result.model_dump()
|
| 58 |
|
| 59 |
|
| 60 |
@app.post("/step")
|
| 61 |
-
def step(req: StepRequest) -> dict:
|
| 62 |
try:
|
| 63 |
-
result =
|
| 64 |
except RuntimeError as e:
|
| 65 |
raise HTTPException(status_code=409, detail=str(e))
|
| 66 |
return result.model_dump()
|
| 67 |
|
| 68 |
|
| 69 |
@app.get("/state")
|
| 70 |
-
def state() -> dict:
|
| 71 |
try:
|
| 72 |
-
return
|
| 73 |
except RuntimeError as e:
|
| 74 |
raise HTTPException(status_code=409, detail=str(e))
|
|
|
|
| 14 |
"""
|
| 15 |
from __future__ import annotations
|
| 16 |
|
| 17 |
+
from uuid import uuid4
|
| 18 |
+
|
| 19 |
from fastapi import FastAPI, HTTPException
|
| 20 |
+
from fastapi import Request, Response
|
| 21 |
|
| 22 |
from src.env import ESCEnv
|
| 23 |
from src.models import ResetRequest, StepRequest
|
|
|
|
| 32 |
),
|
| 33 |
)
|
| 34 |
|
| 35 |
+
SESSION_COOKIE = "esc_session_id"
|
| 36 |
+
_envs: dict[str, ESCEnv] = {}
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def _get_env_for_request(request: Request) -> ESCEnv:
|
| 40 |
+
session_id = request.cookies.get(SESSION_COOKIE)
|
| 41 |
+
if not session_id or session_id not in _envs:
|
| 42 |
+
raise RuntimeError("env.step() called before reset()")
|
| 43 |
+
return _envs[session_id]
|
| 44 |
|
| 45 |
|
| 46 |
@app.get("/")
|
|
|
|
| 59 |
|
| 60 |
|
| 61 |
@app.post("/reset")
|
| 62 |
+
def reset(request: Request, response: Response, req: ResetRequest | None = None) -> dict:
|
| 63 |
req = req or ResetRequest()
|
| 64 |
+
session_id = request.cookies.get(SESSION_COOKIE)
|
| 65 |
+
if not session_id:
|
| 66 |
+
session_id = uuid4().hex
|
| 67 |
+
env = _envs.get(session_id)
|
| 68 |
+
if env is None:
|
| 69 |
+
env = ESCEnv()
|
| 70 |
+
_envs[session_id] = env
|
| 71 |
try:
|
| 72 |
+
result = env.reset(task_id=req.task_id, seed=req.seed)
|
| 73 |
except KeyError as e:
|
| 74 |
raise HTTPException(status_code=400, detail=str(e))
|
| 75 |
+
response.set_cookie(key=SESSION_COOKIE, value=session_id, httponly=True, samesite="lax")
|
| 76 |
return result.model_dump()
|
| 77 |
|
| 78 |
|
| 79 |
@app.post("/step")
|
| 80 |
+
def step(req: StepRequest, request: Request) -> dict:
|
| 81 |
try:
|
| 82 |
+
result = _get_env_for_request(request).step(req.action)
|
| 83 |
except RuntimeError as e:
|
| 84 |
raise HTTPException(status_code=409, detail=str(e))
|
| 85 |
return result.model_dump()
|
| 86 |
|
| 87 |
|
| 88 |
@app.get("/state")
|
| 89 |
+
def state(request: Request) -> dict:
|
| 90 |
try:
|
| 91 |
+
return _get_env_for_request(request).state().model_dump()
|
| 92 |
except RuntimeError as e:
|
| 93 |
raise HTTPException(status_code=409, detail=str(e))
|