Spaces:
Sleeping
Sleeping
Commit ·
57db7b7
1
Parent(s): e849359
Validator 0-1 score change
Browse files- Pre_Validation_Script.sh +0 -185
- Sample_Inference_Script.py +0 -187
- graders.py +16 -8
- test_integration.py +0 -112
- tests_new.py +0 -259
Pre_Validation_Script.sh
DELETED
|
@@ -1,185 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env bash
|
| 2 |
-
#
|
| 3 |
-
# validate-submission.sh — OpenEnv Submission Validator
|
| 4 |
-
#
|
| 5 |
-
# Checks that your HF Space is live, Docker image builds, and openenv validate passes.
|
| 6 |
-
#
|
| 7 |
-
# Prerequisites:
|
| 8 |
-
# - Docker: https://docs.docker.com/get-docker/
|
| 9 |
-
# - openenv-core: pip install openenv-core
|
| 10 |
-
# - curl (usually pre-installed)
|
| 11 |
-
#
|
| 12 |
-
# Run:
|
| 13 |
-
# curl -fsSL https://raw.githubusercontent.com/<owner>/<repo>/main/scripts/validate-submission.sh | bash -s -- <ping_url> [repo_dir]
|
| 14 |
-
#
|
| 15 |
-
# Or download and run locally:
|
| 16 |
-
# chmod +x validate-submission.sh
|
| 17 |
-
# ./validate-submission.sh <ping_url> [repo_dir]
|
| 18 |
-
#
|
| 19 |
-
# Arguments:
|
| 20 |
-
# ping_url Your HuggingFace Space URL (e.g. https://your-space.hf.space)
|
| 21 |
-
# repo_dir Path to your repo (default: current directory)
|
| 22 |
-
#
|
| 23 |
-
# Examples:
|
| 24 |
-
# ./validate-submission.sh https://my-team.hf.space
|
| 25 |
-
# ./validate-submission.sh https://my-team.hf.space ./my-repo
|
| 26 |
-
#
|
| 27 |
-
|
| 28 |
-
set -uo pipefail
|
| 29 |
-
|
| 30 |
-
DOCKER_BUILD_TIMEOUT=600
|
| 31 |
-
if [ -t 1 ]; then
|
| 32 |
-
RED='\033[0;31m'
|
| 33 |
-
GREEN='\033[0;32m'
|
| 34 |
-
YELLOW='\033[1;33m'
|
| 35 |
-
BOLD='\033[1m'
|
| 36 |
-
NC='\033[0m'
|
| 37 |
-
else
|
| 38 |
-
RED='' GREEN='' YELLOW='' BOLD='' NC=''
|
| 39 |
-
fi
|
| 40 |
-
|
| 41 |
-
run_with_timeout() {
|
| 42 |
-
local secs="$1"; shift
|
| 43 |
-
if command -v timeout &>/dev/null; then
|
| 44 |
-
timeout "$secs" "$@"
|
| 45 |
-
elif command -v gtimeout &>/dev/null; then
|
| 46 |
-
gtimeout "$secs" "$@"
|
| 47 |
-
else
|
| 48 |
-
"$@" &
|
| 49 |
-
local pid=$!
|
| 50 |
-
( sleep "$secs" && kill "$pid" 2>/dev/null ) &
|
| 51 |
-
local watcher=$!
|
| 52 |
-
wait "$pid" 2>/dev/null
|
| 53 |
-
local rc=$?
|
| 54 |
-
kill "$watcher" 2>/dev/null
|
| 55 |
-
wait "$watcher" 2>/dev/null
|
| 56 |
-
return $rc
|
| 57 |
-
fi
|
| 58 |
-
}
|
| 59 |
-
|
| 60 |
-
portable_mktemp() {
|
| 61 |
-
local prefix="${1:-validate}"
|
| 62 |
-
mktemp "${TMPDIR:-/tmp}/${prefix}-XXXXXX" 2>/dev/null || mktemp
|
| 63 |
-
}
|
| 64 |
-
|
| 65 |
-
CLEANUP_FILES=()
|
| 66 |
-
cleanup() { rm -f "${CLEANUP_FILES[@]+"${CLEANUP_FILES[@]}"}"; }
|
| 67 |
-
trap cleanup EXIT
|
| 68 |
-
|
| 69 |
-
PING_URL="${1:-}"
|
| 70 |
-
REPO_DIR="${2:-.}"
|
| 71 |
-
|
| 72 |
-
if [ -z "$PING_URL" ]; then
|
| 73 |
-
printf "Usage: %s <ping_url> [repo_dir]\n" "$0"
|
| 74 |
-
printf "\n"
|
| 75 |
-
printf " ping_url Your HuggingFace Space URL (e.g. https://your-space.hf.space)\n"
|
| 76 |
-
printf " repo_dir Path to your repo (default: current directory)\n"
|
| 77 |
-
exit 1
|
| 78 |
-
fi
|
| 79 |
-
|
| 80 |
-
if ! REPO_DIR="$(cd "$REPO_DIR" 2>/dev/null && pwd)"; then
|
| 81 |
-
printf "Error: directory '%s' not found\n" "${2:-.}"
|
| 82 |
-
exit 1
|
| 83 |
-
fi
|
| 84 |
-
PING_URL="${PING_URL%/}"
|
| 85 |
-
export PING_URL
|
| 86 |
-
PASS=0
|
| 87 |
-
|
| 88 |
-
log() { printf "[%s] %b\n" "$(date -u +%H:%M:%S)" "$*"; }
|
| 89 |
-
pass() { log "${GREEN}PASSED${NC} -- $1"; PASS=$((PASS + 1)); }
|
| 90 |
-
fail() { log "${RED}FAILED${NC} -- $1"; }
|
| 91 |
-
hint() { printf " ${YELLOW}Hint:${NC} %b\n" "$1"; }
|
| 92 |
-
stop_at() {
|
| 93 |
-
printf "\n"
|
| 94 |
-
printf "${RED}${BOLD}Validation stopped at %s.${NC} Fix the above before continuing.\n" "$1"
|
| 95 |
-
exit 1
|
| 96 |
-
}
|
| 97 |
-
|
| 98 |
-
printf "\n"
|
| 99 |
-
printf "${BOLD}========================================${NC}\n"
|
| 100 |
-
printf "${BOLD} OpenEnv Submission Validator${NC}\n"
|
| 101 |
-
printf "${BOLD}========================================${NC}\n"
|
| 102 |
-
log "Repo: $REPO_DIR"
|
| 103 |
-
log "Ping URL: $PING_URL"
|
| 104 |
-
printf "\n"
|
| 105 |
-
|
| 106 |
-
log "${BOLD}Step 1/3: Pinging HF Space${NC} ($PING_URL/reset) ..."
|
| 107 |
-
|
| 108 |
-
CURL_OUTPUT=$(portable_mktemp "validate-curl")
|
| 109 |
-
CLEANUP_FILES+=("$CURL_OUTPUT")
|
| 110 |
-
HTTP_CODE=$(curl -s -o "$CURL_OUTPUT" -w "%{http_code}" -X POST \
|
| 111 |
-
-H "Content-Type: application/json" -d '{}' \
|
| 112 |
-
"$PING_URL/reset" --max-time 30 2>"$CURL_OUTPUT" || printf "000")
|
| 113 |
-
|
| 114 |
-
if [ "$HTTP_CODE" = "200" ]; then
|
| 115 |
-
pass "HF Space is live and responds to /reset"
|
| 116 |
-
elif [ "$HTTP_CODE" = "000" ]; then
|
| 117 |
-
fail "HF Space not reachable (connection failed or timed out)"
|
| 118 |
-
hint "Check your network connection and that the Space is running."
|
| 119 |
-
hint "Try: curl -s -o /dev/null -w '%%{http_code}' -X POST $PING_URL/reset"
|
| 120 |
-
stop_at "Step 1"
|
| 121 |
-
else
|
| 122 |
-
fail "HF Space /reset returned HTTP $HTTP_CODE (expected 200)"
|
| 123 |
-
hint "Make sure your Space is running and the URL is correct."
|
| 124 |
-
hint "Try opening $PING_URL in your browser first."
|
| 125 |
-
stop_at "Step 1"
|
| 126 |
-
fi
|
| 127 |
-
|
| 128 |
-
log "${BOLD}Step 2/3: Running docker build${NC} ..."
|
| 129 |
-
|
| 130 |
-
if ! command -v docker &>/dev/null; then
|
| 131 |
-
fail "docker command not found"
|
| 132 |
-
hint "Install Docker: https://docs.docker.com/get-docker/"
|
| 133 |
-
stop_at "Step 2"
|
| 134 |
-
fi
|
| 135 |
-
|
| 136 |
-
if [ -f "$REPO_DIR/Dockerfile" ]; then
|
| 137 |
-
DOCKER_CONTEXT="$REPO_DIR"
|
| 138 |
-
elif [ -f "$REPO_DIR/server/Dockerfile" ]; then
|
| 139 |
-
DOCKER_CONTEXT="$REPO_DIR/server"
|
| 140 |
-
else
|
| 141 |
-
fail "No Dockerfile found in repo root or server/ directory"
|
| 142 |
-
stop_at "Step 2"
|
| 143 |
-
fi
|
| 144 |
-
|
| 145 |
-
log " Found Dockerfile in $DOCKER_CONTEXT"
|
| 146 |
-
|
| 147 |
-
BUILD_OK=false
|
| 148 |
-
BUILD_OUTPUT=$(run_with_timeout "$DOCKER_BUILD_TIMEOUT" docker build "$DOCKER_CONTEXT" 2>&1) && BUILD_OK=true
|
| 149 |
-
|
| 150 |
-
if [ "$BUILD_OK" = true ]; then
|
| 151 |
-
pass "Docker build succeeded"
|
| 152 |
-
else
|
| 153 |
-
fail "Docker build failed (timeout=${DOCKER_BUILD_TIMEOUT}s)"
|
| 154 |
-
printf "%s\n" "$BUILD_OUTPUT" | tail -20
|
| 155 |
-
stop_at "Step 2"
|
| 156 |
-
fi
|
| 157 |
-
|
| 158 |
-
log "${BOLD}Step 3/3: Running openenv validate${NC} ..."
|
| 159 |
-
|
| 160 |
-
if ! command -v openenv &>/dev/null; then
|
| 161 |
-
fail "openenv command not found"
|
| 162 |
-
hint "Install it: pip install openenv-core"
|
| 163 |
-
stop_at "Step 3"
|
| 164 |
-
fi
|
| 165 |
-
|
| 166 |
-
VALIDATE_OK=false
|
| 167 |
-
VALIDATE_OUTPUT=$(cd "$REPO_DIR" && openenv validate 2>&1) && VALIDATE_OK=true
|
| 168 |
-
|
| 169 |
-
if [ "$VALIDATE_OK" = true ]; then
|
| 170 |
-
pass "openenv validate passed"
|
| 171 |
-
[ -n "$VALIDATE_OUTPUT" ] && log " $VALIDATE_OUTPUT"
|
| 172 |
-
else
|
| 173 |
-
fail "openenv validate failed"
|
| 174 |
-
printf "%s\n" "$VALIDATE_OUTPUT"
|
| 175 |
-
stop_at "Step 3"
|
| 176 |
-
fi
|
| 177 |
-
|
| 178 |
-
printf "\n"
|
| 179 |
-
printf "${BOLD}========================================${NC}\n"
|
| 180 |
-
printf "${GREEN}${BOLD} All 3/3 checks passed!${NC}\n"
|
| 181 |
-
printf "${GREEN}${BOLD} Your submission is ready to submit.${NC}\n"
|
| 182 |
-
printf "${BOLD}========================================${NC}\n"
|
| 183 |
-
printf "\n"
|
| 184 |
-
|
| 185 |
-
exit 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Sample_Inference_Script.py
DELETED
|
@@ -1,187 +0,0 @@
|
|
| 1 |
-
"""
|
| 2 |
-
Inference Script Example
|
| 3 |
-
===================================
|
| 4 |
-
MANDATORY
|
| 5 |
-
- Before submitting, ensure the following variables are defined in your environment configuration:
|
| 6 |
-
API_BASE_URL The API endpoint for the LLM.
|
| 7 |
-
MODEL_NAME The model identifier to use for inference.
|
| 8 |
-
HF_TOKEN Your Hugging Face / API key.
|
| 9 |
-
LOCAL_IMAGE_NAME The name of the local image to use for the environment if you are using from_docker_image()
|
| 10 |
-
method
|
| 11 |
-
|
| 12 |
-
- Defaults are set only for API_BASE_URL and MODEL_NAME
|
| 13 |
-
(and should reflect your active inference setup):
|
| 14 |
-
API_BASE_URL = os.getenv("API_BASE_URL", "<your-active-endpoint>")
|
| 15 |
-
MODEL_NAME = os.getenv("MODEL_NAME", "<your-active-model>")
|
| 16 |
-
|
| 17 |
-
- The inference script must be named `inference.py` and placed in the root directory of the project
|
| 18 |
-
- Participants must use OpenAI Client for all LLM calls using above variables
|
| 19 |
-
|
| 20 |
-
STDOUT FORMAT
|
| 21 |
-
- The script must emit exactly three line types to stdout, in this order:
|
| 22 |
-
|
| 23 |
-
[START] task=<task_name> env=<benchmark> model=<model_name>
|
| 24 |
-
[STEP] step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
|
| 25 |
-
[END] success=<true|false> steps=<n> rewards=<r1,r2,...,rn>
|
| 26 |
-
|
| 27 |
-
Rules:
|
| 28 |
-
- One [START] line at episode begin.
|
| 29 |
-
- One [STEP] line per step, immediately after env.step() returns.
|
| 30 |
-
- One [END] line after env.close(), always emitted (even on exception).
|
| 31 |
-
- reward and rewards are formatted to 2 decimal places.
|
| 32 |
-
- done and success are lowercase booleans: true or false.
|
| 33 |
-
- error is the raw last_action_error string, or null if none.
|
| 34 |
-
- All fields on a single line with no newlines within a line.
|
| 35 |
-
|
| 36 |
-
Example:
|
| 37 |
-
[START] task=click-test env=miniwob model=Qwen3-VL-30B
|
| 38 |
-
[STEP] step=1 action=click('123') reward=0.00 done=false error=null
|
| 39 |
-
[STEP] step=2 action=fill('456','text') reward=0.00 done=false error=null
|
| 40 |
-
[STEP] step=3 action=click('789') reward=1.00 done=true error=null
|
| 41 |
-
[END] success=true steps=3 rewards=0.00,0.00,1.00
|
| 42 |
-
"""
|
| 43 |
-
|
| 44 |
-
import asyncio
|
| 45 |
-
import os
|
| 46 |
-
import textwrap
|
| 47 |
-
from typing import List, Optional
|
| 48 |
-
|
| 49 |
-
from openai import OpenAI
|
| 50 |
-
|
| 51 |
-
from my_env_v4 import MyEnvV4Action, MyEnvV4Env
|
| 52 |
-
IMAGE_NAME = os.getenv("IMAGE_NAME") # If you are using docker image
|
| 53 |
-
API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
|
| 54 |
-
|
| 55 |
-
API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"
|
| 56 |
-
MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct"
|
| 57 |
-
TASK_NAME = os.getenv("MY_ENV_V4_TASK", "echo")
|
| 58 |
-
BENCHMARK = os.getenv("MY_ENV_V4_BENCHMARK", "my_env_v4")
|
| 59 |
-
MAX_STEPS = 8
|
| 60 |
-
TEMPERATURE = 0.7
|
| 61 |
-
MAX_TOKENS = 150
|
| 62 |
-
SUCCESS_SCORE_THRESHOLD = 0.1 # normalized score in [0, 1]
|
| 63 |
-
|
| 64 |
-
# Max possible reward: each token contributes 0.1, across all steps
|
| 65 |
-
_MAX_REWARD_PER_STEP = MAX_TOKENS * 0.1
|
| 66 |
-
MAX_TOTAL_REWARD = MAX_STEPS * _MAX_REWARD_PER_STEP
|
| 67 |
-
|
| 68 |
-
SYSTEM_PROMPT = textwrap.dedent(
|
| 69 |
-
"""
|
| 70 |
-
You are interacting with a simple echo environment.
|
| 71 |
-
Each turn you must send a message. The environment will echo it back.
|
| 72 |
-
Reward is proportional to message length: reward = len(message) * 0.1
|
| 73 |
-
Your goal is to maximize total reward by sending meaningful, substantive messages.
|
| 74 |
-
Reply with exactly one message string — no quotes, no prefixes, just the message text.
|
| 75 |
-
"""
|
| 76 |
-
).strip()
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
def log_start(task: str, env: str, model: str) -> None:
|
| 80 |
-
print(f"[START] task={task} env={env} model={model}", flush=True)
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
|
| 84 |
-
error_val = error if error else "null"
|
| 85 |
-
done_val = str(done).lower()
|
| 86 |
-
print(
|
| 87 |
-
f"[STEP] step={step} action={action} reward={reward:.2f} done={done_val} error={error_val}",
|
| 88 |
-
flush=True,
|
| 89 |
-
)
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
|
| 93 |
-
rewards_str = ",".join(f"{r:.2f}" for r in rewards)
|
| 94 |
-
print(f"[END] success={str(success).lower()} steps={steps} score={score:.3f} rewards={rewards_str}", flush=True)
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
def build_user_prompt(step: int, last_echoed: str, last_reward: float, history: List[str]) -> str:
|
| 98 |
-
history_block = "\n".join(history[-4:]) if history else "None"
|
| 99 |
-
return textwrap.dedent(
|
| 100 |
-
f"""
|
| 101 |
-
Step: {step}
|
| 102 |
-
Last echoed message: {last_echoed!r}
|
| 103 |
-
Last reward: {last_reward:.2f}
|
| 104 |
-
Previous steps:
|
| 105 |
-
{history_block}
|
| 106 |
-
Send your next message.
|
| 107 |
-
"""
|
| 108 |
-
).strip()
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
def get_model_message(client: OpenAI, step: int, last_echoed: str, last_reward: float, history: List[str]) -> str:
|
| 112 |
-
user_prompt = build_user_prompt(step, last_echoed, last_reward, history)
|
| 113 |
-
try:
|
| 114 |
-
completion = client.chat.completions.create(
|
| 115 |
-
model=MODEL_NAME,
|
| 116 |
-
messages=[
|
| 117 |
-
{"role": "system", "content": SYSTEM_PROMPT},
|
| 118 |
-
{"role": "user", "content": user_prompt},
|
| 119 |
-
],
|
| 120 |
-
temperature=TEMPERATURE,
|
| 121 |
-
max_tokens=MAX_TOKENS,
|
| 122 |
-
stream=False,
|
| 123 |
-
)
|
| 124 |
-
text = (completion.choices[0].message.content or "").strip()
|
| 125 |
-
return text if text else "hello"
|
| 126 |
-
except Exception as exc:
|
| 127 |
-
print(f"[DEBUG] Model request failed: {exc}", flush=True)
|
| 128 |
-
return "hello"
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
async def main() -> None:
|
| 132 |
-
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 133 |
-
|
| 134 |
-
env = await MyEnvV4Env.from_docker_image(IMAGE_NAME)
|
| 135 |
-
|
| 136 |
-
history: List[str] = []
|
| 137 |
-
rewards: List[float] = []
|
| 138 |
-
steps_taken = 0
|
| 139 |
-
score = 0.0
|
| 140 |
-
success = False
|
| 141 |
-
|
| 142 |
-
log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
|
| 143 |
-
|
| 144 |
-
try:
|
| 145 |
-
result = await env.reset() # OpenENV.reset()
|
| 146 |
-
last_echoed = result.observation.echoed_message
|
| 147 |
-
last_reward = 0.0
|
| 148 |
-
|
| 149 |
-
for step in range(1, MAX_STEPS + 1):
|
| 150 |
-
if result.done:
|
| 151 |
-
break
|
| 152 |
-
|
| 153 |
-
message = get_model_message(client, step, last_echoed, last_reward, history)
|
| 154 |
-
|
| 155 |
-
result = await env.step(MyEnvV4Action(message=message))
|
| 156 |
-
obs = result.observation
|
| 157 |
-
|
| 158 |
-
reward = result.reward or 0.0
|
| 159 |
-
done = result.done
|
| 160 |
-
error = None
|
| 161 |
-
|
| 162 |
-
rewards.append(reward)
|
| 163 |
-
steps_taken = step
|
| 164 |
-
last_echoed = obs.echoed_message
|
| 165 |
-
last_reward = reward
|
| 166 |
-
|
| 167 |
-
log_step(step=step, action=message, reward=reward, done=done, error=error)
|
| 168 |
-
|
| 169 |
-
history.append(f"Step {step}: {message!r} -> reward {reward:+.2f}")
|
| 170 |
-
|
| 171 |
-
if done:
|
| 172 |
-
break
|
| 173 |
-
|
| 174 |
-
score = sum(rewards) / MAX_TOTAL_REWARD if MAX_TOTAL_REWARD > 0 else 0.0
|
| 175 |
-
score = min(max(score, 0.0), 1.0) # clamp to [0, 1]
|
| 176 |
-
success = score >= SUCCESS_SCORE_THRESHOLD
|
| 177 |
-
|
| 178 |
-
finally:
|
| 179 |
-
try:
|
| 180 |
-
await env.close()
|
| 181 |
-
except Exception as e:
|
| 182 |
-
print(f"[DEBUG] env.close() error (container cleanup): {e}", flush=True)
|
| 183 |
-
log_end(success=success, steps=steps_taken, score=score, rewards=rewards)
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
if __name__ == "__main__":
|
| 187 |
-
asyncio.run(main())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
graders.py
CHANGED
|
@@ -2,7 +2,7 @@
|
|
| 2 |
Deterministic graders for SupportEnv tasks.
|
| 3 |
|
| 4 |
Each grader inspects the agent's action_history against ground-truth data
|
| 5 |
-
and returns (score, breakdown, feedback) where score is in
|
| 6 |
|
| 7 |
Task 1 — Classification: category match (0.50) + priority match (0.40) + efficiency (0.10)
|
| 8 |
Task 2 — Extraction: entity coverage (0.60) + action coverage (0.30) + no hallucination (0.10)
|
|
@@ -14,6 +14,14 @@ from __future__ import annotations
|
|
| 14 |
from typing import Any, Dict, List, Optional, Tuple
|
| 15 |
|
| 16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
def grade_task(
|
| 18 |
task_id: str, episode_state: Dict[str, Any]
|
| 19 |
) -> Tuple[float, Dict[str, float], str]:
|
|
@@ -23,7 +31,7 @@ def grade_task(
|
|
| 23 |
return _grade_extraction(episode_state)
|
| 24 |
elif task_id == "task3":
|
| 25 |
return _grade_resolution(episode_state)
|
| 26 |
-
return 0.0, {}, "Unknown task"
|
| 27 |
|
| 28 |
|
| 29 |
# ---------------------------------------------------------------------------
|
|
@@ -66,7 +74,7 @@ def _grade_classification(ep: Dict[str, Any]) -> Tuple[float, Dict[str, float],
|
|
| 66 |
|
| 67 |
classify_action = _last_action_of_type(history, "classify")
|
| 68 |
if classify_action is None:
|
| 69 |
-
return 0.0, breakdown, "No classify action found."
|
| 70 |
|
| 71 |
# Category
|
| 72 |
if _normalize(classify_action.get("category")) == _normalize(gt["category"]):
|
|
@@ -84,7 +92,7 @@ def _grade_classification(ep: Dict[str, Any]) -> Tuple[float, Dict[str, float],
|
|
| 84 |
else:
|
| 85 |
breakdown["efficiency"] = round(max(0.0, 0.10 * (1 - (steps_used - 1) / max_steps)), 4)
|
| 86 |
|
| 87 |
-
score =
|
| 88 |
parts = ", ".join(f"{k}={v:.2f}" for k, v in breakdown.items())
|
| 89 |
return score, breakdown, f"Task 1: {parts}"
|
| 90 |
|
|
@@ -112,7 +120,7 @@ def _grade_extraction(ep: Dict[str, Any]) -> Tuple[float, Dict[str, float], str]
|
|
| 112 |
extract_action = _last_action_of_type(history, "extract")
|
| 113 |
if extract_action is None:
|
| 114 |
breakdown["no_hallucination"] = 0.0
|
| 115 |
-
return 0.0, breakdown, "No extract action found."
|
| 116 |
|
| 117 |
# --- Entity coverage ---
|
| 118 |
gt_entities: Dict[str, Any] = gt.get("entities", {})
|
|
@@ -144,7 +152,7 @@ def _grade_extraction(ep: Dict[str, Any]) -> Tuple[float, Dict[str, float], str]
|
|
| 144 |
penalty = min(len(extra_keys) * 0.02, 0.10)
|
| 145 |
breakdown["no_hallucination"] = round(max(0.0, 0.10 - penalty), 4)
|
| 146 |
|
| 147 |
-
score =
|
| 148 |
parts = ", ".join(f"{k}={v:.2f}" for k, v in breakdown.items())
|
| 149 |
return score, breakdown, f"Task 2: {parts}"
|
| 150 |
|
|
@@ -185,7 +193,7 @@ def _grade_resolution(ep: Dict[str, Any]) -> Tuple[float, Dict[str, float], str]
|
|
| 185 |
respond_action = _last_action_of_type(history, "respond")
|
| 186 |
if respond_action is None:
|
| 187 |
breakdown["no_empty_steps"] = 0.0
|
| 188 |
-
return 0.0, breakdown, "No respond action found."
|
| 189 |
|
| 190 |
response_text: str = respond_action.get("response_text") or ""
|
| 191 |
resolution_steps: List[str] = respond_action.get("resolution_steps") or []
|
|
@@ -239,6 +247,6 @@ def _grade_resolution(ep: Dict[str, Any]) -> Tuple[float, Dict[str, float], str]
|
|
| 239 |
if not resolution_steps or any(not s.strip() for s in resolution_steps):
|
| 240 |
breakdown["no_empty_steps"] = 0.0
|
| 241 |
|
| 242 |
-
score =
|
| 243 |
parts = ", ".join(f"{k}={v:.2f}" for k, v in breakdown.items())
|
| 244 |
return score, breakdown, f"Task 3: {parts}"
|
|
|
|
| 2 |
Deterministic graders for SupportEnv tasks.
|
| 3 |
|
| 4 |
Each grader inspects the agent's action_history against ground-truth data
|
| 5 |
+
and returns (score, breakdown, feedback) where score is in (0.0, 1.0).
|
| 6 |
|
| 7 |
Task 1 — Classification: category match (0.50) + priority match (0.40) + efficiency (0.10)
|
| 8 |
Task 2 — Extraction: entity coverage (0.60) + action coverage (0.30) + no hallucination (0.10)
|
|
|
|
| 14 |
from typing import Any, Dict, List, Optional, Tuple
|
| 15 |
|
| 16 |
|
| 17 |
+
SCORE_EPSILON = 0.0001
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def _strict_score(score: float) -> float:
|
| 21 |
+
"""Map any score into the strict open interval (0, 1)."""
|
| 22 |
+
return round(min(max(score, SCORE_EPSILON), 1.0 - SCORE_EPSILON), 4)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
def grade_task(
|
| 26 |
task_id: str, episode_state: Dict[str, Any]
|
| 27 |
) -> Tuple[float, Dict[str, float], str]:
|
|
|
|
| 31 |
return _grade_extraction(episode_state)
|
| 32 |
elif task_id == "task3":
|
| 33 |
return _grade_resolution(episode_state)
|
| 34 |
+
return _strict_score(0.0), {}, "Unknown task"
|
| 35 |
|
| 36 |
|
| 37 |
# ---------------------------------------------------------------------------
|
|
|
|
| 74 |
|
| 75 |
classify_action = _last_action_of_type(history, "classify")
|
| 76 |
if classify_action is None:
|
| 77 |
+
return _strict_score(0.0), breakdown, "No classify action found."
|
| 78 |
|
| 79 |
# Category
|
| 80 |
if _normalize(classify_action.get("category")) == _normalize(gt["category"]):
|
|
|
|
| 92 |
else:
|
| 93 |
breakdown["efficiency"] = round(max(0.0, 0.10 * (1 - (steps_used - 1) / max_steps)), 4)
|
| 94 |
|
| 95 |
+
score = _strict_score(sum(breakdown.values()))
|
| 96 |
parts = ", ".join(f"{k}={v:.2f}" for k, v in breakdown.items())
|
| 97 |
return score, breakdown, f"Task 1: {parts}"
|
| 98 |
|
|
|
|
| 120 |
extract_action = _last_action_of_type(history, "extract")
|
| 121 |
if extract_action is None:
|
| 122 |
breakdown["no_hallucination"] = 0.0
|
| 123 |
+
return _strict_score(0.0), breakdown, "No extract action found."
|
| 124 |
|
| 125 |
# --- Entity coverage ---
|
| 126 |
gt_entities: Dict[str, Any] = gt.get("entities", {})
|
|
|
|
| 152 |
penalty = min(len(extra_keys) * 0.02, 0.10)
|
| 153 |
breakdown["no_hallucination"] = round(max(0.0, 0.10 - penalty), 4)
|
| 154 |
|
| 155 |
+
score = _strict_score(sum(breakdown.values()))
|
| 156 |
parts = ", ".join(f"{k}={v:.2f}" for k, v in breakdown.items())
|
| 157 |
return score, breakdown, f"Task 2: {parts}"
|
| 158 |
|
|
|
|
| 193 |
respond_action = _last_action_of_type(history, "respond")
|
| 194 |
if respond_action is None:
|
| 195 |
breakdown["no_empty_steps"] = 0.0
|
| 196 |
+
return _strict_score(0.0), breakdown, "No respond action found."
|
| 197 |
|
| 198 |
response_text: str = respond_action.get("response_text") or ""
|
| 199 |
resolution_steps: List[str] = respond_action.get("resolution_steps") or []
|
|
|
|
| 247 |
if not resolution_steps or any(not s.strip() for s in resolution_steps):
|
| 248 |
breakdown["no_empty_steps"] = 0.0
|
| 249 |
|
| 250 |
+
score = _strict_score(sum(breakdown.values()))
|
| 251 |
parts = ", ".join(f"{k}={v:.2f}" for k, v in breakdown.items())
|
| 252 |
return score, breakdown, f"Task 3: {parts}"
|
test_integration.py
DELETED
|
@@ -1,112 +0,0 @@
|
|
| 1 |
-
"""
|
| 2 |
-
Integration test for SupportEnv.
|
| 3 |
-
|
| 4 |
-
Runs a full episode for each task and prints results.
|
| 5 |
-
Usage: python test_integration.py
|
| 6 |
-
"""
|
| 7 |
-
import environment as env
|
| 8 |
-
from models import Action
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
def test_task(task_id: str) -> bool:
|
| 12 |
-
"""Run a full episode for a task. Returns True if passed."""
|
| 13 |
-
print(f"\n{'='*60}")
|
| 14 |
-
print(f"Testing {task_id}")
|
| 15 |
-
print(f"{'='*60}")
|
| 16 |
-
|
| 17 |
-
# Reset
|
| 18 |
-
print("1. reset()...")
|
| 19 |
-
obs = env.reset(task_id, ticket_index=0)
|
| 20 |
-
episode_id = obs.episode_id
|
| 21 |
-
print(f" [OK] episode_id={episode_id[:8]}...")
|
| 22 |
-
print(f" ticket_id={obs.ticket.ticket_id} subject={obs.ticket.subject[:50]}")
|
| 23 |
-
print(f" max_steps={obs.max_steps} hint={obs.hint}")
|
| 24 |
-
|
| 25 |
-
# Take a relevant action
|
| 26 |
-
print("2. step() with task action...")
|
| 27 |
-
if task_id == "task1":
|
| 28 |
-
action = Action(action_type="classify", category="billing", priority="high")
|
| 29 |
-
elif task_id == "task2":
|
| 30 |
-
action = Action(
|
| 31 |
-
action_type="extract",
|
| 32 |
-
extracted_entities={"customer_name": "Robert Chen", "account_id": "ACC-78234"},
|
| 33 |
-
required_actions=["issue_refund"],
|
| 34 |
-
)
|
| 35 |
-
else: # task3
|
| 36 |
-
action = Action(
|
| 37 |
-
action_type="respond",
|
| 38 |
-
response_text=(
|
| 39 |
-
"We sincerely apologize for the inconvenience with your password reset. "
|
| 40 |
-
"We will manually reset your password and send a new email immediately. "
|
| 41 |
-
"Please check your spam folder and whitelist our domain. "
|
| 42 |
-
"We will resolve this within the next 30 minutes."
|
| 43 |
-
),
|
| 44 |
-
resolution_steps=[
|
| 45 |
-
"verify_email_delivery",
|
| 46 |
-
"check_spam_filters",
|
| 47 |
-
"manual_password_reset",
|
| 48 |
-
"follow_up_confirmation",
|
| 49 |
-
],
|
| 50 |
-
)
|
| 51 |
-
|
| 52 |
-
result = env.step(episode_id, action)
|
| 53 |
-
print(f" [OK] step_reward={result.reward.step_reward:+.4f} done={result.done}")
|
| 54 |
-
|
| 55 |
-
# Submit
|
| 56 |
-
print("3. step() submit...")
|
| 57 |
-
if not result.done:
|
| 58 |
-
result = env.step(episode_id, Action(action_type="submit"))
|
| 59 |
-
print(f" [OK] done={result.done} total_reward={result.reward.total_reward:.4f}")
|
| 60 |
-
|
| 61 |
-
# State
|
| 62 |
-
print("4. get_state()...")
|
| 63 |
-
state = env.get_state(episode_id)
|
| 64 |
-
print(f" [OK] steps={state.step_number} history_len={len(state.history)}")
|
| 65 |
-
|
| 66 |
-
# Grade
|
| 67 |
-
print("5. grade()...")
|
| 68 |
-
score, breakdown, feedback = env.grade(episode_id)
|
| 69 |
-
print(f" [OK] score={score:.4f}/1.0")
|
| 70 |
-
print(f" breakdown: {', '.join(f'{k}={v:.2f}' for k, v in breakdown.items())}")
|
| 71 |
-
print(f" feedback: {feedback}")
|
| 72 |
-
|
| 73 |
-
passed = score >= 0.0 # just verify pipeline works
|
| 74 |
-
return passed
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
def main():
|
| 78 |
-
print("SupportEnv Integration Test")
|
| 79 |
-
print("=" * 60)
|
| 80 |
-
|
| 81 |
-
results = []
|
| 82 |
-
for task_id in ["task1", "task2", "task3"]:
|
| 83 |
-
try:
|
| 84 |
-
ok = test_task(task_id)
|
| 85 |
-
results.append((task_id, ok, None))
|
| 86 |
-
except Exception as exc:
|
| 87 |
-
import traceback
|
| 88 |
-
traceback.print_exc()
|
| 89 |
-
results.append((task_id, False, str(exc)))
|
| 90 |
-
finally:
|
| 91 |
-
env._EPISODES.clear()
|
| 92 |
-
|
| 93 |
-
print(f"\n{'='*60}")
|
| 94 |
-
print("SUMMARY")
|
| 95 |
-
print("=" * 60)
|
| 96 |
-
all_ok = True
|
| 97 |
-
for task_id, ok, err in results:
|
| 98 |
-
status = "[PASS]" if ok else "[FAIL]"
|
| 99 |
-
print(f" {status} {task_id}" + (f" — {err}" if err else ""))
|
| 100 |
-
if not ok:
|
| 101 |
-
all_ok = False
|
| 102 |
-
|
| 103 |
-
if all_ok:
|
| 104 |
-
print("\n[OK] All integration tests passed!")
|
| 105 |
-
else:
|
| 106 |
-
print("\n[FAIL] Some tests failed.")
|
| 107 |
-
return 0 if all_ok else 1
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
if __name__ == "__main__":
|
| 111 |
-
import sys
|
| 112 |
-
sys.exit(main())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
tests_new.py
DELETED
|
@@ -1,259 +0,0 @@
|
|
| 1 |
-
"""
|
| 2 |
-
Comprehensive tests for SupportEnv.
|
| 3 |
-
|
| 4 |
-
Run with: pytest tests_new.py -v
|
| 5 |
-
"""
|
| 6 |
-
import pytest
|
| 7 |
-
|
| 8 |
-
import environment as env
|
| 9 |
-
from models import Action, Observation, StepResult, State
|
| 10 |
-
from data import TASK_META
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
class TestReset:
|
| 14 |
-
"""Test episode reset functionality."""
|
| 15 |
-
|
| 16 |
-
def test_reset_task1(self):
|
| 17 |
-
obs = env.reset("task1")
|
| 18 |
-
assert isinstance(obs, Observation)
|
| 19 |
-
assert obs.task_id == "task1"
|
| 20 |
-
assert obs.episode_id is not None
|
| 21 |
-
assert obs.ticket is not None
|
| 22 |
-
assert obs.step_number == 0
|
| 23 |
-
assert obs.max_steps == TASK_META["task1"]["max_steps"]
|
| 24 |
-
|
| 25 |
-
def test_reset_task2(self):
|
| 26 |
-
obs = env.reset("task2")
|
| 27 |
-
assert obs.task_id == "task2"
|
| 28 |
-
assert obs.ticket.ticket_id.startswith("T2-")
|
| 29 |
-
|
| 30 |
-
def test_reset_task3(self):
|
| 31 |
-
obs = env.reset("task3")
|
| 32 |
-
assert obs.task_id == "task3"
|
| 33 |
-
assert obs.ticket.ticket_id.startswith("T3-")
|
| 34 |
-
|
| 35 |
-
def test_reset_with_ticket_index(self):
|
| 36 |
-
obs = env.reset("task1", ticket_index=2)
|
| 37 |
-
assert obs.ticket.ticket_id == "T1-003"
|
| 38 |
-
|
| 39 |
-
def test_reset_invalid_task(self):
|
| 40 |
-
with pytest.raises(ValueError):
|
| 41 |
-
env.reset("task_unknown")
|
| 42 |
-
|
| 43 |
-
def test_reset_invalid_ticket_index(self):
|
| 44 |
-
with pytest.raises(ValueError):
|
| 45 |
-
env.reset("task1", ticket_index=99)
|
| 46 |
-
|
| 47 |
-
def test_reset_creates_episode(self):
|
| 48 |
-
obs = env.reset("task1")
|
| 49 |
-
assert obs.episode_id in env._EPISODES
|
| 50 |
-
ep = env._EPISODES[obs.episode_id]
|
| 51 |
-
assert ep["task_id"] == "task1"
|
| 52 |
-
assert ep["step_number"] == 0
|
| 53 |
-
assert ep["done"] is False
|
| 54 |
-
|
| 55 |
-
def test_reset_hint_on_first_step(self):
|
| 56 |
-
obs = env.reset("task1")
|
| 57 |
-
assert obs.hint is not None
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
class TestStep:
|
| 61 |
-
"""Test step execution."""
|
| 62 |
-
|
| 63 |
-
def test_step_classify(self):
|
| 64 |
-
obs = env.reset("task1")
|
| 65 |
-
action = Action(action_type="classify", category="billing", priority="high")
|
| 66 |
-
result = env.step(obs.episode_id, action)
|
| 67 |
-
assert isinstance(result, StepResult)
|
| 68 |
-
assert result.observation.step_number == 1
|
| 69 |
-
assert result.reward.step_reward is not None
|
| 70 |
-
|
| 71 |
-
def test_step_extract(self):
|
| 72 |
-
obs = env.reset("task2")
|
| 73 |
-
action = Action(
|
| 74 |
-
action_type="extract",
|
| 75 |
-
extracted_entities={"customer_name": "Alice"},
|
| 76 |
-
required_actions=["issue_refund"],
|
| 77 |
-
)
|
| 78 |
-
result = env.step(obs.episode_id, action)
|
| 79 |
-
assert result.observation.step_number == 1
|
| 80 |
-
|
| 81 |
-
def test_step_respond(self):
|
| 82 |
-
obs = env.reset("task3")
|
| 83 |
-
action = Action(
|
| 84 |
-
action_type="respond",
|
| 85 |
-
response_text="Thank you for reaching out. We sincerely apologize for the inconvenience and will resolve this immediately.",
|
| 86 |
-
resolution_steps=["verify_account", "issue_refund"],
|
| 87 |
-
)
|
| 88 |
-
result = env.step(obs.episode_id, action)
|
| 89 |
-
assert result.observation.step_number == 1
|
| 90 |
-
|
| 91 |
-
def test_step_submit_marks_done(self):
|
| 92 |
-
obs = env.reset("task1")
|
| 93 |
-
result = env.step(obs.episode_id, Action(action_type="submit"))
|
| 94 |
-
assert result.done is True
|
| 95 |
-
assert result.observation.available_actions == []
|
| 96 |
-
|
| 97 |
-
def test_step_invalid_episode(self):
|
| 98 |
-
with pytest.raises(KeyError):
|
| 99 |
-
env.step("nonexistent-id", Action(action_type="submit"))
|
| 100 |
-
|
| 101 |
-
def test_step_after_done_raises(self):
|
| 102 |
-
obs = env.reset("task1")
|
| 103 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 104 |
-
with pytest.raises(ValueError):
|
| 105 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 106 |
-
|
| 107 |
-
def test_step_max_steps_ends_episode(self):
|
| 108 |
-
obs = env.reset("task1")
|
| 109 |
-
max_steps = obs.max_steps
|
| 110 |
-
for i in range(max_steps):
|
| 111 |
-
action = Action(action_type="classify", category="general", priority="low")
|
| 112 |
-
result = env.step(obs.episode_id, action)
|
| 113 |
-
assert result.done is True
|
| 114 |
-
|
| 115 |
-
def test_thread_history_grows(self):
|
| 116 |
-
obs = env.reset("task1")
|
| 117 |
-
env.step(obs.episode_id, Action(action_type="classify", category="billing", priority="high"))
|
| 118 |
-
result = env.step(obs.episode_id, Action(action_type="submit"))
|
| 119 |
-
assert len(result.observation.thread_history) == 2
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
class TestState:
|
| 123 |
-
"""Test state retrieval."""
|
| 124 |
-
|
| 125 |
-
def test_get_state_initial(self):
|
| 126 |
-
obs = env.reset("task1")
|
| 127 |
-
state = env.get_state(obs.episode_id)
|
| 128 |
-
assert isinstance(state, State)
|
| 129 |
-
assert state.episode_id == obs.episode_id
|
| 130 |
-
assert state.task_id == "task1"
|
| 131 |
-
assert state.step_number == 0
|
| 132 |
-
assert state.done is False
|
| 133 |
-
|
| 134 |
-
def test_get_state_invalid(self):
|
| 135 |
-
with pytest.raises(KeyError):
|
| 136 |
-
env.get_state("bad-id")
|
| 137 |
-
|
| 138 |
-
def test_state_history_after_step(self):
|
| 139 |
-
obs = env.reset("task2")
|
| 140 |
-
env.step(obs.episode_id, Action(action_type="extract", extracted_entities={}, required_actions=[]))
|
| 141 |
-
state = env.get_state(obs.episode_id)
|
| 142 |
-
assert len(state.history) == 1
|
| 143 |
-
assert state.history[0]["action_type"] == "extract"
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
class TestGraders:
|
| 147 |
-
"""Test grading for each task."""
|
| 148 |
-
|
| 149 |
-
def test_grade_task1_perfect(self):
|
| 150 |
-
"""Correct category + priority on ticket 0 (billing/high)."""
|
| 151 |
-
obs = env.reset("task1", ticket_index=0)
|
| 152 |
-
env.step(obs.episode_id, Action(action_type="classify", category="billing", priority="high"))
|
| 153 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 154 |
-
score, breakdown, feedback = env.grade(obs.episode_id)
|
| 155 |
-
assert score >= 0.9
|
| 156 |
-
assert breakdown["category_correct"] == 0.50
|
| 157 |
-
assert breakdown["priority_correct"] == 0.40
|
| 158 |
-
|
| 159 |
-
def test_grade_task1_wrong_category(self):
|
| 160 |
-
obs = env.reset("task1", ticket_index=0)
|
| 161 |
-
env.step(obs.episode_id, Action(action_type="classify", category="technical", priority="high"))
|
| 162 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 163 |
-
score, breakdown, _ = env.grade(obs.episode_id)
|
| 164 |
-
assert breakdown["category_correct"] == 0.0
|
| 165 |
-
assert breakdown["priority_correct"] == 0.40
|
| 166 |
-
|
| 167 |
-
def test_grade_task1_no_classify_action(self):
|
| 168 |
-
obs = env.reset("task1")
|
| 169 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 170 |
-
score, _, _ = env.grade(obs.episode_id)
|
| 171 |
-
assert score == 0.0
|
| 172 |
-
|
| 173 |
-
def test_grade_task2_entities(self):
|
| 174 |
-
obs = env.reset("task2", ticket_index=0)
|
| 175 |
-
env.step(obs.episode_id, Action(
|
| 176 |
-
action_type="extract",
|
| 177 |
-
extracted_entities={
|
| 178 |
-
"customer_name": "Robert Chen",
|
| 179 |
-
"account_id": "ACC-78234",
|
| 180 |
-
"invoice_number": "INV-20240312",
|
| 181 |
-
"incorrect_amount": "199.00",
|
| 182 |
-
"correct_amount": "99.00",
|
| 183 |
-
"refund_amount": "100.00",
|
| 184 |
-
},
|
| 185 |
-
required_actions=["issue_refund", "send_corrected_invoice"],
|
| 186 |
-
))
|
| 187 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 188 |
-
score, breakdown, _ = env.grade(obs.episode_id)
|
| 189 |
-
assert breakdown["entity_coverage"] == pytest.approx(0.60, abs=0.01)
|
| 190 |
-
assert breakdown["action_coverage"] == pytest.approx(0.30, abs=0.01)
|
| 191 |
-
|
| 192 |
-
def test_grade_task3_keywords_and_steps(self):
|
| 193 |
-
obs = env.reset("task3", ticket_index=0)
|
| 194 |
-
env.step(obs.episode_id, Action(
|
| 195 |
-
action_type="respond",
|
| 196 |
-
response_text=(
|
| 197 |
-
"We sincerely apologize for the password reset issue. "
|
| 198 |
-
"We will send a new reset email and ask you to check your spam folder "
|
| 199 |
-
"and whitelist our domain. We will have this resolved within the hour."
|
| 200 |
-
),
|
| 201 |
-
resolution_steps=[
|
| 202 |
-
"verify_email_delivery",
|
| 203 |
-
"check_spam_filters",
|
| 204 |
-
"manual_password_reset",
|
| 205 |
-
"follow_up_confirmation",
|
| 206 |
-
],
|
| 207 |
-
))
|
| 208 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 209 |
-
score, breakdown, _ = env.grade(obs.episode_id)
|
| 210 |
-
assert score >= 0.7
|
| 211 |
-
assert breakdown["length_adequate"] == 0.10
|
| 212 |
-
assert breakdown["no_empty_steps"] == 0.05
|
| 213 |
-
|
| 214 |
-
def test_grade_not_done_raises(self):
|
| 215 |
-
obs = env.reset("task1")
|
| 216 |
-
with pytest.raises(ValueError):
|
| 217 |
-
env.grade(obs.episode_id)
|
| 218 |
-
|
| 219 |
-
def test_grade_invalid_episode_raises(self):
|
| 220 |
-
with pytest.raises(KeyError):
|
| 221 |
-
env.grade("bad-id")
|
| 222 |
-
|
| 223 |
-
def test_score_in_range(self):
|
| 224 |
-
for task_id in ["task1", "task2", "task3"]:
|
| 225 |
-
obs = env.reset(task_id)
|
| 226 |
-
env.step(obs.episode_id, Action(action_type="submit"))
|
| 227 |
-
score, _, _ = env.grade(obs.episode_id)
|
| 228 |
-
assert 0.0 <= score <= 1.0
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
class TestRewards:
|
| 232 |
-
"""Test reward signals."""
|
| 233 |
-
|
| 234 |
-
def test_step_reward_is_float(self):
|
| 235 |
-
obs = env.reset("task1")
|
| 236 |
-
result = env.step(obs.episode_id, Action(action_type="classify", category="billing", priority="high"))
|
| 237 |
-
assert isinstance(result.reward.step_reward, float)
|
| 238 |
-
|
| 239 |
-
def test_total_reward_accumulates(self):
|
| 240 |
-
obs = env.reset("task2")
|
| 241 |
-
r1 = env.step(obs.episode_id, Action(action_type="extract", extracted_entities={}, required_actions=[]))
|
| 242 |
-
r2 = env.step(obs.episode_id, Action(action_type="submit"))
|
| 243 |
-
assert r2.reward.total_reward != r1.reward.total_reward
|
| 244 |
-
|
| 245 |
-
def test_submit_bonus_applied(self):
|
| 246 |
-
obs = env.reset("task1")
|
| 247 |
-
result = env.step(obs.episode_id, Action(action_type="submit"))
|
| 248 |
-
# submit_bonus=0.05 minus step_cost=0.02 = +0.03 base before grader
|
| 249 |
-
assert result.reward.step_reward > 0.0
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
@pytest.fixture(autouse=True)
|
| 253 |
-
def cleanup():
|
| 254 |
-
yield
|
| 255 |
-
env._EPISODES.clear()
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
if __name__ == "__main__":
|
| 259 |
-
pytest.main([__file__, "-v"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|