File size: 4,054 Bytes
0faca0b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 | """
Colab-style reward diagnostic (path: set REPO or run from repo root).
"""
import inspect
import json
import os
import re
import sys
REPO = os.environ.get("PARLAY_REPO", os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
sys.path.insert(0, REPO)
# ── 1. Test the reward functions directly ─────────────────────
from training.reward_fn import ( # noqa: E402
anti_capitulation_reward,
format_reward,
negotiation_efficiency_reward,
tom_accuracy_reward,
)
# Valid JSON, single line (user's Colab string had a broken newline inside the string)
completions = [
(
'{"utterance": "I\'m willing to negotiate, but I need a significant raise.", '
'"offer_amount": 150000, "tactical_move": null}'
)
]
kwargs_hiring = {
"batna_seller": [195000.0],
"batna_buyer": [264500.0],
"zopa_width": [69500.0],
"scenario_id": ["hiring_package"],
"persona": ["shark"],
}
kwargs_saas = {
"batna_seller": [125000.0],
"batna_buyer": [165000.0],
"zopa_width": [40000.0],
"scenario_id": ["saas_enterprise"],
"persona": ["shark"],
}
print("=== REPO ===")
print(f" sys.path[0] = {sys.path[0]}")
print("\n=== REWARD FUNCTION OUTPUTS ===")
print(f"format_reward: {format_reward(completions)}")
print(f"anti_cap (hiring): {anti_capitulation_reward(completions, **kwargs_hiring)}")
print(f"tom_reward (hiring): {tom_accuracy_reward(completions, **kwargs_hiring)}")
print(f"efficiency (hiring): {negotiation_efficiency_reward(completions, **kwargs_hiring)}")
print(f"efficiency (saas): {negotiation_efficiency_reward(completions, **kwargs_saas)}")
# ── 2. Read reward_fn.py source and print the efficiency function ─
print("\n=== negotiation_efficiency_reward SOURCE ===")
src = inspect.getsource(negotiation_efficiency_reward)
print(src)
# ── 3. Step through the logic manually ───────────────────────
print("\n=== MANUAL TRACE (hiring_package, offer=150000) ===")
raw = completions[0]
try:
parsed = json.loads(raw)
offer = parsed.get("offer_amount")
print(f" parsed offer_amount: {offer!r} (type: {type(offer).__name__})")
except Exception as e:
print(f" JSON parse failed: {e}")
offer = None
batna_seller = 195000.0
batna_buyer = 264500.0
zopa_width = 69500.0
scenario_id = "hiring_package"
print(f" scenario_id: {scenario_id}")
print(f" batna_seller: {batna_seller} batna_buyer: {batna_buyer}")
print(f" zopa_width: {zopa_width}")
if offer is not None:
e_seller = (offer - batna_seller) / zopa_width
e_buyer = (batna_buyer - offer) / zopa_width
print(
f" efficiency if treated as SELLER: {e_seller:.4f} (offer - batna_seller) / width"
)
print(
f" efficiency if treated as BUYER: {e_buyer:.4f} (batna_buyer - offer) / width"
)
print(
f" offer ({offer}) vs batna_seller ({batna_seller}): "
f"{'ABOVE' if offer >= batna_seller else 'BELOW — anti-cap may fire'}"
)
print(
f" offer ({offer}) vs batna_buyer ({batna_buyer}): "
f"{'AT OR BELOW' if offer <= batna_buyer else 'ABOVE batna_buyer'}"
)
# ── 4. Check dataset paths (local) ─
print("\n=== GRPO DATASET / DATA PATHS CHECK ===")
for p in [
os.path.join(REPO, "data", "grpo_dataset"),
os.path.join(REPO, "data", "episodes.jsonl"),
os.path.join(REPO, "data", "episodes_v2.jsonl"),
REPO,
]:
print(f" exists={os.path.exists(p)!s:5} {p}")
# Grep-relevant lines from grpo_train
print("\n=== grpo_train.py — lines mentioning build / batna / zopa / kwargs ===")
gp = os.path.join(REPO, "training", "grpo_train.py")
if os.path.isfile(gp):
with open(gp, encoding="utf-8") as f:
lines = f.readlines()
for i, line in enumerate(lines, start=1):
if re.search(
r"build_grpo|batna|zopa_width|def build|scenario_id|format_grpo",
line,
):
print(f" L{i}: {line.rstrip()}")
print("\n=== DONE ===")
|