Commit: c85f6b41fc7a2feabf601e41537d7cefa0f4c4fa - zirpu (git)

Emergent generative agents

Commit MetaInfo

• Revision: c85f6b41fc7a2feabf601e41537d7cefa0f4c4fa (tree)
• Zeit: 2023-04-10 03:57:49
• Autor: Corbin <cds@corb...>
• Commiter: Corbin

Log Message

IRC stuff.

Multiple channels, timestamps, better formatting of log lines, log
storage, and small tweaks.

Ändern Zusammenfassung

• modified: agent.py
• modified: common.py
• modified: gens/camelid.py

Diff

--- a/agent.py
+++ b/agent.py
@@ -1,7 +1,11 @@
 #!/usr/bin/env nix-shell
 #! nix-shell -i python3 -p python3Packages.irc python3Packages.transformers python3Packages.torch
 
+from concurrent.futures import ThreadPoolExecutor
+from datetime import datetime
 import json
+import os.path
+import random
 import sys
 
 from irc.bot import SingleServerIRCBot
@@ -52,58 +56,95 @@ Goals, working memory, notes: {allGoals}
 def load_character(path):
     with open(path) as handle: return json.load(handle)
 
-characters = [load_character(arg) for arg in sys.argv[1:]]
-prompts = [build_prompt(**character) for character in characters]
-prompt, title = prompts.pop(0)
+character = load_character(sys.argv[2])
+startingChannels = character.pop("startingChannels")
+prompt, title = build_prompt(**character)
 
 MAX_NEW_TOKENS = 128
-gen = CamelidGen()
-gen = HFGen(Flavor.GPTNeo, MAX_NEW_TOKENS)
+if sys.argv[1] == "llama":
+    print("~ Initializing camelid adapter…")
+    gen = CamelidGen()
+else:
+    print("~ Initializing GPT-Neo on HF…")
+    gen = HFGen(Flavor.GPTNeo, MAX_NEW_TOKENS)
 max_context_length = gen.contextLength()
 
+executor = ThreadPoolExecutor(max_workers=1)
+
 class Agent(SingleServerIRCBot):
-    def __init__(self, host, title):
+    def __init__(self, host, title, startingChannels, logpath):
         super(Agent, self).__init__([(host, 6667)], title_to_nick(title), title)
-        self.log = Log([])
+        self.startingChannels = startingChannels
+        self.logpath = logpath
+        self.logs = {}
+
+    def on_join(self, c, e):
+        channel = e.target
+        c.topic(channel)
+        try:
+            with open(os.path.join(self.logpath, channel + ".txt"), "r") as f:
+                self.logs[channel] = Log(f.read().strip().split("\n"))
+        except IOError: self.logs[channel] = Log([])
 
-    def on_join(self, c, e): c.topic(e.target)
     def on_currenttopic(self, c, e):
         self.channels[e.arguments[0]].topic = e.arguments[1]
 
-    def on_welcome(self, c, e): c.join("#treehouse")
+    def on_welcome(self, c, e):
+        for channel in self.startingChannels: c.join(channel)
 
     def on_pubmsg(self, c, e):
         line = e.arguments[0]
-        self.log.push(e.source.nick, line)
+        channel = e.target
+        log = self.logs[channel]
+        log.irc(datetime.now(), e.source.nick, line)
         # Dispatch in the style of
         # https://github.com/jaraco/irc/blob/main/scripts/testbot.py
-        if ":" in line:
-            prefix = lower(self.connection.get_nickname()) + ":"
-            if lower(line).startswith(prefix):
-                self.do_command(c, e, line[len(prefix):].strip())
-
-    def do_command(self, c, e, inst):
-        channel = e.target
-        nick = self.connection.get_nickname()
-        users = self.channels[channel].users()
-        prefix = nick + ":"
-        fullPrompt = prompt + self.chatPrompt(channel)
-        self.log.bumpCutoff(max_context_length, gen.countTokens, fullPrompt, prefix)
-        s = self.log.finishPrompt(fullPrompt, prefix)
-        print("~ log cutoff:", self.log.cutoff,
-              "prompt length (tokens):", gen.countTokens(s))
-        # Hack: GPT-Neo tries to smuggle messages, so forbid it.
-        forbidden = [prefix] + ["." + user for user in users]
-        reply = parseLine(gen.complete(s), forbidden)
-        self.log.push(nick, reply)
-        c.privmsg(channel, reply)
+        nick = lower(self.connection.get_nickname())
+        colon = nick + ":"
+        comma = nick + ","
+        lowered = lower(line)
+        if colon in lowered or comma in lowered:
+            self.do_command(c, e, line[len(colon):].strip())
+        elif nick in lowered and random.random() <= 0.875:
+            self.generateReply(c, channel)
+        elif random.random() <= 0.125: self.generateReply(c, channel)
+
+    def do_command(self, c, e, inst): self.generateReply(c, e.target)
 
     def chatPrompt(self, channel):
-        topic = getattr(self.channels[channel], "topic", "no topic")
+        c = self.channels[channel]
+        topic = getattr(c, "topic", "no topic")
+        users = ", ".join(c.users())
         return f"""
 IRC channel: {channel}
 Channel topic: {topic}
+Channel users: {users}
 """
 
-agent = Agent("june.local", title)
-agent.start()
+    def generateReply(self, c, channel):
+        log = self.logs[channel]
+        nick = self.connection.get_nickname()
+        users = self.channels[channel].users()
+        prefix = nick + ":"
+        fullPrompt = prompt + self.chatPrompt(channel)
+        log.bumpCutoff(max_context_length, gen.countTokens, fullPrompt, prefix)
+        s = log.finishPrompt(fullPrompt, prefix)
+        print("~ log cutoff:", log.cutoff,
+              "prompt length (tokens):", gen.countTokens(s))
+        forbidden = [prefix] + ["." + user for user in users]
+        # NB: At this point, execution is kicked out to a thread.
+        def cb(completion):
+            reply = parseLine(completion.result(), forbidden)
+            log.irc(datetime.now(), nick, reply)
+            c.privmsg(channel, reply)
+        executor.submit(lambda: gen.complete(s)).add_done_callback(cb)
+
+logpath = sys.argv[3]
+agent = Agent("june.local", title, startingChannels, logpath)
+try: agent.start()
+except KeyboardInterrupt:
+    print("~ Saving logs…")
+    for channel, log in agent.logs.items():
+        with open(os.path.join(agent.logpath, channel + ".txt"), "w") as f:
+            f.write("\n".join(log.l))
+print("~ Quitting, bye!")

--- a/common.py
+++ b/common.py
@@ -22,6 +22,8 @@ class Log:
         self.stamp += 1
 
     def push(self, speaker, entry): self.raw(speaker + ": " + entry)
+    def irc(self, t, speaker, entry):
+        self.raw(f"{t:%H:%M:%S} <{speaker}> {entry}")
 
     def finishPrompt(self, s, prefix):
         return self.finishPromptAtCutoff(self.cutoff, s, prefix)

--- a/gens/camelid.py
+++ b/gens/camelid.py
@@ -4,9 +4,10 @@ from common import Timer
 
 llama = [
     "/home/simpson/models/result/bin/llama-cpp",
-     "-m", "/home/simpson/models/13b-4bit.bin",
-     "-f", "/dev/stdin",
-     "-c", "2048",
+    "-t", "3",
+    "-m", "/home/simpson/models/7b-4bit.bin",
+    "-f", "/dev/stdin",
+    "-c", "2048",
 ]
 
 class CamelidGen: