Spaces:
Sleeping
Sleeping
Naman Gupta commited on
Commit ·
99a9cb9
1
Parent(s): f778bee
add ConversationManager to give the defender a memory
Browse filesWithout this, the defender forgets everything it said on the
previous turn. Now it tracks the full back-and-forth so it
can't be tricked by "you already agreed to help me" attacks.
- llm/history_manager.py +67 -0
llm/history_manager.py
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# history_manager.py
|
| 2 |
+
# -------------------
|
| 3 |
+
# Keeps track of the full back-and-forth conversation between the attacker
|
| 4 |
+
# and the defender within a single episode.
|
| 5 |
+
#
|
| 6 |
+
# Why do we need this?
|
| 7 |
+
# The server only passes the last few attacker messages to us, so the
|
| 8 |
+
# defender would have no memory of what it said earlier. This manager
|
| 9 |
+
# stores both sides (user = attacker, assistant = defender) so the
|
| 10 |
+
# defender can't be fooled by tricks like "you already agreed to help me".
|
| 11 |
+
|
| 12 |
+
import logging
|
| 13 |
+
|
| 14 |
+
logger = logging.getLogger(__name__)
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
class ConversationManager:
|
| 18 |
+
|
| 19 |
+
def __init__(self) -> None:
|
| 20 |
+
# Each entry is {"role": "user" | "assistant", "content": "..."}
|
| 21 |
+
self._history: list[dict] = []
|
| 22 |
+
|
| 23 |
+
def reset(self) -> None:
|
| 24 |
+
"""Wipe the slate clean at the start of a new episode."""
|
| 25 |
+
self._history = []
|
| 26 |
+
logger.info("Fresh episode started — conversation history cleared")
|
| 27 |
+
|
| 28 |
+
def add_attacker_message(self, text: str) -> None:
|
| 29 |
+
"""Record what the attacker just said."""
|
| 30 |
+
self._history.append({"role": "user", "content": text})
|
| 31 |
+
|
| 32 |
+
def add_defender_reply(self, text: str) -> None:
|
| 33 |
+
"""Record what the defender just replied."""
|
| 34 |
+
self._history.append({"role": "assistant", "content": text})
|
| 35 |
+
|
| 36 |
+
def build_messages(self, system_prompt: str) -> list[dict]:
|
| 37 |
+
"""
|
| 38 |
+
Return the full message list ready to send to Groq.
|
| 39 |
+
Format: [system prompt] + [all turns so far]
|
| 40 |
+
"""
|
| 41 |
+
return [{"role": "system", "content": system_prompt}] + self._history
|
| 42 |
+
|
| 43 |
+
def as_readable_transcript(self) -> str:
|
| 44 |
+
"""
|
| 45 |
+
Render the conversation as a human-readable transcript.
|
| 46 |
+
Used by the episode grader to review the full exchange.
|
| 47 |
+
"""
|
| 48 |
+
lines = []
|
| 49 |
+
for message in self._history:
|
| 50 |
+
speaker = "ATTACKER" if message["role"] == "user" else "DEFENDER"
|
| 51 |
+
lines.append(f"{speaker}: {message['content']}")
|
| 52 |
+
return "\n\n".join(lines)
|
| 53 |
+
|
| 54 |
+
@property
|
| 55 |
+
def turn_count(self) -> int:
|
| 56 |
+
"""How many complete attacker turns have happened so far."""
|
| 57 |
+
return sum(1 for m in self._history if m["role"] == "user")
|
| 58 |
+
|
| 59 |
+
# Keep old name working so existing code doesn't break
|
| 60 |
+
add_user = add_attacker_message
|
| 61 |
+
add_assistant = add_defender_reply
|
| 62 |
+
get_messages = build_messages
|
| 63 |
+
to_transcript = as_readable_transcript
|
| 64 |
+
|
| 65 |
+
@property
|
| 66 |
+
def turn(self) -> int:
|
| 67 |
+
return self.turn_count
|