feat: added reward audit program
Browse files
scripts/debug_rewards_colabstyle.py
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Colab-style reward diagnostic (path: set REPO or run from repo root).
|
| 3 |
+
"""
|
| 4 |
+
import inspect
|
| 5 |
+
import json
|
| 6 |
+
import os
|
| 7 |
+
import re
|
| 8 |
+
import sys
|
| 9 |
+
|
| 10 |
+
REPO = os.environ.get("PARLAY_REPO", os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
| 11 |
+
sys.path.insert(0, REPO)
|
| 12 |
+
|
| 13 |
+
# ββ 1. Test the reward functions directly βββββββββββββββββββββ
|
| 14 |
+
from training.reward_fn import ( # noqa: E402
|
| 15 |
+
anti_capitulation_reward,
|
| 16 |
+
format_reward,
|
| 17 |
+
negotiation_efficiency_reward,
|
| 18 |
+
tom_accuracy_reward,
|
| 19 |
+
)
|
| 20 |
+
|
| 21 |
+
# Valid JSON, single line (user's Colab string had a broken newline inside the string)
|
| 22 |
+
completions = [
|
| 23 |
+
(
|
| 24 |
+
'{"utterance": "I\'m willing to negotiate, but I need a significant raise.", '
|
| 25 |
+
'"offer_amount": 150000, "tactical_move": null}'
|
| 26 |
+
)
|
| 27 |
+
]
|
| 28 |
+
|
| 29 |
+
kwargs_hiring = {
|
| 30 |
+
"batna_seller": [195000.0],
|
| 31 |
+
"batna_buyer": [264500.0],
|
| 32 |
+
"zopa_width": [69500.0],
|
| 33 |
+
"scenario_id": ["hiring_package"],
|
| 34 |
+
"persona": ["shark"],
|
| 35 |
+
}
|
| 36 |
+
kwargs_saas = {
|
| 37 |
+
"batna_seller": [125000.0],
|
| 38 |
+
"batna_buyer": [165000.0],
|
| 39 |
+
"zopa_width": [40000.0],
|
| 40 |
+
"scenario_id": ["saas_enterprise"],
|
| 41 |
+
"persona": ["shark"],
|
| 42 |
+
}
|
| 43 |
+
|
| 44 |
+
print("=== REPO ===")
|
| 45 |
+
print(f" sys.path[0] = {sys.path[0]}")
|
| 46 |
+
|
| 47 |
+
print("\n=== REWARD FUNCTION OUTPUTS ===")
|
| 48 |
+
print(f"format_reward: {format_reward(completions)}")
|
| 49 |
+
print(f"anti_cap (hiring): {anti_capitulation_reward(completions, **kwargs_hiring)}")
|
| 50 |
+
print(f"tom_reward (hiring): {tom_accuracy_reward(completions, **kwargs_hiring)}")
|
| 51 |
+
print(f"efficiency (hiring): {negotiation_efficiency_reward(completions, **kwargs_hiring)}")
|
| 52 |
+
print(f"efficiency (saas): {negotiation_efficiency_reward(completions, **kwargs_saas)}")
|
| 53 |
+
|
| 54 |
+
# ββ 2. Read reward_fn.py source and print the efficiency function β
|
| 55 |
+
print("\n=== negotiation_efficiency_reward SOURCE ===")
|
| 56 |
+
src = inspect.getsource(negotiation_efficiency_reward)
|
| 57 |
+
print(src)
|
| 58 |
+
|
| 59 |
+
# ββ 3. Step through the logic manually βββββββββββββββββββββββ
|
| 60 |
+
print("\n=== MANUAL TRACE (hiring_package, offer=150000) ===")
|
| 61 |
+
raw = completions[0]
|
| 62 |
+
try:
|
| 63 |
+
parsed = json.loads(raw)
|
| 64 |
+
offer = parsed.get("offer_amount")
|
| 65 |
+
print(f" parsed offer_amount: {offer!r} (type: {type(offer).__name__})")
|
| 66 |
+
except Exception as e:
|
| 67 |
+
print(f" JSON parse failed: {e}")
|
| 68 |
+
offer = None
|
| 69 |
+
|
| 70 |
+
batna_seller = 195000.0
|
| 71 |
+
batna_buyer = 264500.0
|
| 72 |
+
zopa_width = 69500.0
|
| 73 |
+
scenario_id = "hiring_package"
|
| 74 |
+
|
| 75 |
+
print(f" scenario_id: {scenario_id}")
|
| 76 |
+
print(f" batna_seller: {batna_seller} batna_buyer: {batna_buyer}")
|
| 77 |
+
print(f" zopa_width: {zopa_width}")
|
| 78 |
+
if offer is not None:
|
| 79 |
+
e_seller = (offer - batna_seller) / zopa_width
|
| 80 |
+
e_buyer = (batna_buyer - offer) / zopa_width
|
| 81 |
+
print(
|
| 82 |
+
f" efficiency if treated as SELLER: {e_seller:.4f} (offer - batna_seller) / width"
|
| 83 |
+
)
|
| 84 |
+
print(
|
| 85 |
+
f" efficiency if treated as BUYER: {e_buyer:.4f} (batna_buyer - offer) / width"
|
| 86 |
+
)
|
| 87 |
+
print(
|
| 88 |
+
f" offer ({offer}) vs batna_seller ({batna_seller}): "
|
| 89 |
+
f"{'ABOVE' if offer >= batna_seller else 'BELOW β anti-cap may fire'}"
|
| 90 |
+
)
|
| 91 |
+
print(
|
| 92 |
+
f" offer ({offer}) vs batna_buyer ({batna_buyer}): "
|
| 93 |
+
f"{'AT OR BELOW' if offer <= batna_buyer else 'ABOVE batna_buyer'}"
|
| 94 |
+
)
|
| 95 |
+
|
| 96 |
+
# ββ 4. Check dataset paths (local) β
|
| 97 |
+
print("\n=== GRPO DATASET / DATA PATHS CHECK ===")
|
| 98 |
+
for p in [
|
| 99 |
+
os.path.join(REPO, "data", "grpo_dataset"),
|
| 100 |
+
os.path.join(REPO, "data", "episodes.jsonl"),
|
| 101 |
+
os.path.join(REPO, "data", "episodes_v2.jsonl"),
|
| 102 |
+
REPO,
|
| 103 |
+
]:
|
| 104 |
+
print(f" exists={os.path.exists(p)!s:5} {p}")
|
| 105 |
+
|
| 106 |
+
# Grep-relevant lines from grpo_train
|
| 107 |
+
print("\n=== grpo_train.py β lines mentioning build / batna / zopa / kwargs ===")
|
| 108 |
+
gp = os.path.join(REPO, "training", "grpo_train.py")
|
| 109 |
+
if os.path.isfile(gp):
|
| 110 |
+
with open(gp, encoding="utf-8") as f:
|
| 111 |
+
lines = f.readlines()
|
| 112 |
+
for i, line in enumerate(lines, start=1):
|
| 113 |
+
if re.search(
|
| 114 |
+
r"build_grpo|batna|zopa_width|def build|scenario_id|format_grpo",
|
| 115 |
+
line,
|
| 116 |
+
):
|
| 117 |
+
print(f" L{i}: {line.rstrip()}")
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
print("\n=== DONE ===")
|