kossy/Bernie
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
The Bernie AI chat bot. It feels the burn.
5 commits 1 branch 0 tags 5 MiB
  • Python 100%
Find a file
Exact
kossy 89bc1b5977
 don't even know
2026-06-17 11:06:09 -07:00
assets took a new pic of Bernie 2026-06-13 15:43:36 -07:00
docker reorg docker 2026-06-13 19:41:15 -07:00
Bernie.py init 2026-06-13 15:37:04 -07:00
CHANGELOG.md init 2026-06-13 15:37:04 -07:00
README.md added some docker gayness 2026-06-13 16:05:12 -07:00
TODO.md don't even know 2026-06-17 11:06:09 -07:00

README.md

image

Bernie chats bot

An IRC bot that learns from IRC logs and generates NPC-style messages using Markov chains, Ollama, LM Studio, or a local CUDA/Hugging Face Transformers model.

The bot can optionally support:

  • !say
  • !rss <feed-url>
  • RSS/Atom/Git feed announcements
  • periodic NPC-style speech
  • live learning from channel messages
  • random automatic replies
  • Shabbat quiet mode
  • local CUDA model generation

The !bernie command has been removed.

Table of Contents

Features

  • Import IRC logs from a file or directory
  • Build and save a reusable Markov brain
  • Generate fictional NPC-style IRC messages
  • Use one of four brain backends:
    • markov
    • ollama
    • lmstudio
    • localcuda
  • Use local LLMs through:
    • Ollama
    • LM Studio
    • PyTorch/Hugging Face Transformers
  • Fall back to Markov generation if an LLM backend fails
  • Add RSS/Atom/Git feeds at runtime
  • Periodically announce RSS feed updates
  • Optionally learn from live channel messages
  • Optionally generate random replies
  • Optionally suppress generated speech during Shabbat

Installation

Basic dependencies

pip install requests feedparser

Make the bot executable:

chmod +x irc_smart_log_bot.py

Optional local CUDA dependencies

Only needed if using:

--brain-backend localcuda

Install:

pip install torch transformers accelerate

For the correct CUDA-specific PyTorch command, check:

https://pytorch.org/get-started/locally/

Optional Ollama setup

Install Ollama:

https://ollama.com/

Start Ollama:

ollama serve

Pull a model:

ollama pull llama3.1:8b

Optional LM Studio setup

Install LM Studio:

https://lmstudio.ai/

Then:

  1. Load a model.
  2. Start the local server.
  3. Confirm the API base URL.

Default local API URL:

http://127.0.0.1:1234/v1

Quick Start

Example Markov bot with !say enabled:

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend markov \
  --enable-say-command

IRC Commands

Commands are typed directly in IRC.

!say

Generate one NPC/brain line.

Usage

!say

Requires

--enable-say-command

Example response

[rkl-npc] yeah that sounds about right

If --enable-say-command is not enabled, the bot ignores !say.

!rss <feed-url>

Add an RSS, Atom, or Git feed at runtime.

Usage

!rss <feed-url>

Example

!rss https://github.com/example/project/commits/main.atom

Requires

--enable-rss --enable-rss-command

If RSS polling is not enabled, the bot will not add or announce feeds.

Removed Commands

!bernie

The !bernie command has been removed.

There is no built-in IRC help command by default. Use this README as the command reference.

Command-Line Options

IRC Connection Options

Option Description Example
--server IRC server hostname --server irc.libera.chat
--port IRC server port --port 6697
--tls Enable TLS/SSL --tls
--nick Bot nickname --nick StyleBot
--channel Channel to join --channel '#test'
--password Optional server password --password secret
--username IRC username --username stylebot
--realname IRC real name field --realname "Style Bot"

Log and Brain Options

Option Description Default
--logs File or directory containing IRC logs ./logs
--target-nick Nickname to learn from in logs none
--brain Saved Markov brain file brain.pkl
--rebuild-brain Rebuild the brain from logs disabled
--markov-order Markov chain order 2

Example:

--logs ./logs --target-nick rkl --brain brain.pkl --rebuild-brain

NPC Output Options

Option Description Default
--npc-name Fictional NPC name used in prompts npc
--npc-prefix Prefix added to generated messages [npc]

Example:

--npc-name rkl --npc-prefix "[rkl-npc] "

Using an NPC prefix is recommended so generated messages are not confused with real users.

Feature Flags

Option Description
--enable-say-command Enables the !say IRC command
--enable-rss Enables RSS/Atom/Git feed polling
--enable-rss-command Enables the !rss <feed-url> IRC command
--enable-periodic-brain Enables automatic periodic NPC messages
--enable-live-learning Learns from non-command channel messages while running
--enable-random-replies Allows occasional automatic replies
--enable-shabbat Suppresses generated speech during Shabbat quiet mode

Random Reply Options

Option Description Default
--random-reply-chance Chance of replying to a normal message 0.02

Example:

--enable-random-replies --random-reply-chance 0.05

RSS Options

Option Description Default
--rss-feed Add an RSS/Atom/Git feed at startup. Can be used multiple times. none
--rss-interval RSS polling interval in seconds 300

Example:

--enable-rss \
--rss-feed https://github.com/example/project/commits/main.atom \
--rss-interval 300

Periodic NPC Options

Option Description Default
--periodic-min-seconds Minimum delay between periodic messages 300
--periodic-max-seconds Maximum delay between periodic messages 1200

Example:

--enable-periodic-brain \
--periodic-min-seconds 600 \
--periodic-max-seconds 1800

Brain Backends

The bot supports four brain backends:

--brain-backend markov
--brain-backend ollama
--brain-backend lmstudio
--brain-backend localcuda
Backend Description
markov Uses a local Markov chain trained from IRC logs
ollama Uses a local Ollama HTTP API model
lmstudio Uses LM Studio's OpenAI-compatible local API
localcuda Uses PyTorch/Hugging Face Transformers directly

Markov Backend

The Markov backend is the default backend.

It imports IRC logs and builds a Markov chain.

Usage

--brain-backend markov

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend markov \
  --enable-say-command

Ollama Backend

The Ollama backend uses a local Ollama model through the Ollama HTTP API.

Requirements

Install Ollama:

https://ollama.com/

Start Ollama:

ollama serve

Pull a model:

ollama pull llama3.1:8b

Usage

--brain-backend ollama

Options

Option Description Default
--ollama-host Ollama API host http://127.0.0.1:11434
--ollama-model Ollama model name llama3.1:8b

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend ollama \
  --ollama-host http://127.0.0.1:11434 \
  --ollama-model llama3.1:8b \
  --llm-fallback-markov \
  --enable-say-command

LM Studio Backend

The LM Studio backend uses LM Studio's OpenAI-compatible local API.

Requirements

Install LM Studio:

https://lmstudio.ai/

Then:

  1. Load a model in LM Studio.
  2. Start the local server.
  3. Confirm the local server URL.

Default:

http://127.0.0.1:1234/v1

Usage

--brain-backend lmstudio

Options

Option Description Default
--lmstudio-base-url LM Studio OpenAI-compatible API base URL http://127.0.0.1:1234/v1
--lmstudio-model LM Studio model name local-model

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend lmstudio \
  --lmstudio-base-url http://127.0.0.1:1234/v1 \
  --lmstudio-model local-model \
  --llm-fallback-markov \
  --enable-say-command

Local CUDA Backend

The local CUDA backend uses PyTorch and Hugging Face Transformers directly.

This is useful if Ollama or LM Studio are not available.

Requirements

Install dependencies:

pip install torch transformers accelerate

For NVIDIA CUDA, use the correct PyTorch installation command from:

https://pytorch.org/get-started/locally/

Usage

--brain-backend localcuda

Options

Option Description Default
--localcuda-model Hugging Face model name or local model path mistralai/Mistral-7B-Instruct-v0.2
--localcuda-device Device to use: cuda or cpu cuda

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick CudaBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend localcuda \
  --localcuda-model mistralai/Mistral-7B-Instruct-v0.2 \
  --localcuda-device cuda \
  --llm-fallback-markov \
  --enable-say-command

CPU mode

You can force CPU mode:

--brain-backend localcuda --localcuda-device cpu

CPU mode may be very slow for large models.

LLM Options

These options apply to:

  • ollama
  • lmstudio
  • localcuda
Option Description Default
--llm-max-new-tokens Maximum generated tokens 120
--llm-temperature Randomness of generated output 0.8
--llm-top-p Nucleus sampling value 0.95
--llm-fallback-markov Fall back to Markov if LLM generation fails disabled

Example:

--llm-max-new-tokens 120 \
--llm-temperature 0.8 \
--llm-top-p 0.95 \
--llm-fallback-markov

RSS Usage

Add feeds at startup

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick FeedBot \
  --channel '#test' \
  --enable-rss \
  --rss-feed https://github.com/example/project/commits/main.atom

Add feeds from IRC

Start with:

--enable-rss --enable-rss-command

Then type in IRC:

!rss https://github.com/example/project/commits/main.atom

The bot announces new entries like:

[rss] Commit title https://github.com/example/project/commit/abc123

Examples

Example 1: Simple Markov bot with !say

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend markov \
  --enable-say-command

Example 2: Ollama bot with Markov fallback

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend ollama \
  --ollama-model llama3.1:8b \
  --llm-fallback-markov \
  --enable-say-command

Example 3: LM Studio with RSS

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StudioBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend lmstudio \
  --lmstudio-base-url http://127.0.0.1:1234/v1 \
  --lmstudio-model local-model \
  --llm-fallback-markov \
  --enable-say-command \
  --enable-rss \
  --enable-rss-command

Example 4: Local CUDA with Markov fallback

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick CudaBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend localcuda \
  --localcuda-model mistralai/Mistral-7B-Instruct-v0.2 \
  --localcuda-device cuda \
  --llm-fallback-markov \
  --enable-say-command

Example 5: Periodic NPC messages

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick NPCBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --enable-periodic-brain \
  --periodic-min-seconds 600 \
  --periodic-max-seconds 1800

Example 6: Live learning and random replies

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick ChatBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --enable-live-learning \
  --enable-random-replies \
  --random-reply-chance 0.03

Example 7: RSS feed bot

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick FeedBot \
  --channel '#test' \
  --enable-rss \
  --enable-rss-command \
  --rss-feed https://github.com/example/project/commits/main.atom \
  --rss-interval 300

Example 8: Shabbat quiet mode

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick QuietBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --enable-say-command \
  --enable-periodic-brain \
  --enable-random-replies \
  --enable-shabbat

Troubleshooting

!say does nothing

Make sure the bot was started with:

--enable-say-command

!rss does nothing

Make sure the bot was started with:

--enable-rss --enable-rss-command

Also install:

pip install feedparser

Ollama generation fails

Make sure Ollama is running:

ollama serve

Check installed models:

ollama list

Pull a model if needed:

ollama pull llama3.1:8b

LM Studio generation fails

Check that:

  1. A model is loaded.
  2. The local server is running.
  3. The base URL is correct.

Default:

http://127.0.0.1:1234/v1

Local CUDA generation fails

Install dependencies:

pip install torch transformers accelerate

Check CUDA availability:

python3 -c "import torch; print(torch.cuda.is_available())"

If CUDA is not available, either fix your PyTorch/CUDA install or use:

--localcuda-device cpu

CPU mode may be very slow for large models.

Model download fails

Some Hugging Face models require accepting a license or logging in.

Install the Hugging Face CLI:

pip install huggingface_hub

Login:

huggingface-cli login

Then retry.

Out of memory with local CUDA

Try:

  • a smaller model
  • a quantized model
  • lower --llm-max-new-tokens
  • closing other GPU applications
  • using Ollama or LM Studio instead

Example smaller model:

--localcuda-model gpt2

Markov output is empty

Check that:

  • --logs points to the correct file or directory
  • --target-nick matches the nick in your logs
  • the logs contain normal IRC messages
  • the brain was rebuilt after changing logs

Rebuild with:

--rebuild-brain

RSS feed updates repeat after restart

The bot tracks seen feed entries in memory.

If the bot restarts, it may announce recent feed entries again.

Security and Identity Notes

Generated messages should be treated as fictional NPC-style output.

Recommended:

--npc-prefix "[npc] "

or:

--npc-prefix "[rkl-npc] "

This helps users distinguish generated text from real user messages.

Do not use this bot to impersonate real users.

CUDA and GPU Notes

CUDA/GPU can be used in three ways:

  1. Through Ollama
  2. Through LM Studio
  3. Directly through --brain-backend localcuda

When using Ollama or LM Studio, GPU configuration is handled by those tools.

When using localcuda, GPU configuration is handled by PyTorch and Transformers.

image

Bernie chats bot

An IRC bot that learns from IRC logs and generates NPC-style messages using Markov chains, Ollama, LM Studio, or a local CUDA/Hugging Face Transformers model.

The bot can optionally support:

  • !say
  • !rss <feed-url>
  • RSS/Atom/Git feed announcements
  • periodic NPC-style speech
  • live learning from channel messages
  • random automatic replies
  • Shabbat quiet mode
  • local CUDA model generation

The !bernie command has been removed.

Table of Contents

Features

  • Import IRC logs from a file or directory
  • Build and save a reusable Markov brain
  • Generate fictional NPC-style IRC messages
  • Use one of four brain backends:
    • markov
    • ollama
    • lmstudio
    • localcuda
  • Use local LLMs through:
    • Ollama
    • LM Studio
    • PyTorch/Hugging Face Transformers
  • Fall back to Markov generation if an LLM backend fails
  • Add RSS/Atom/Git feeds at runtime
  • Periodically announce RSS feed updates
  • Optionally learn from live channel messages
  • Optionally generate random replies
  • Optionally suppress generated speech during Shabbat

Installation

Basic dependencies

pip install requests feedparser

Make the bot executable:

chmod +x irc_smart_log_bot.py

Optional local CUDA dependencies

Only needed if using:

--brain-backend localcuda

Install:

pip install torch transformers accelerate

For the correct CUDA-specific PyTorch command, check:

https://pytorch.org/get-started/locally/

Optional Ollama setup

Install Ollama:

https://ollama.com/

Start Ollama:

ollama serve

Pull a model:

ollama pull llama3.1:8b

Optional LM Studio setup

Install LM Studio:

https://lmstudio.ai/

Then:

  1. Load a model.
  2. Start the local server.
  3. Confirm the API base URL.

Default local API URL:

http://127.0.0.1:1234/v1

Quick Start

Example Markov bot with !say enabled:

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend markov \
  --enable-say-command

IRC Commands

Commands are typed directly in IRC.

!say

Generate one NPC/brain line.

Usage

!say

Requires

--enable-say-command

Example response

[rkl-npc] yeah that sounds about right

If --enable-say-command is not enabled, the bot ignores !say.

!rss <feed-url>

Add an RSS, Atom, or Git feed at runtime.

Usage

!rss <feed-url>

Example

!rss https://github.com/example/project/commits/main.atom

Requires

--enable-rss --enable-rss-command

If RSS polling is not enabled, the bot will not add or announce feeds.

Removed Commands

!bernie

The !bernie command has been removed.

There is no built-in IRC help command by default. Use this README as the command reference.

Command-Line Options

IRC Connection Options

Option Description Example
--server IRC server hostname --server irc.libera.chat
--port IRC server port --port 6697
--tls Enable TLS/SSL --tls
--nick Bot nickname --nick StyleBot
--channel Channel to join --channel '#test'
--password Optional server password --password secret
--username IRC username --username stylebot
--realname IRC real name field --realname "Style Bot"

Log and Brain Options

Option Description Default
--logs File or directory containing IRC logs ./logs
--target-nick Nickname to learn from in logs none
--brain Saved Markov brain file brain.pkl
--rebuild-brain Rebuild the brain from logs disabled
--markov-order Markov chain order 2

Example:

--logs ./logs --target-nick rkl --brain brain.pkl --rebuild-brain

NPC Output Options

Option Description Default
--npc-name Fictional NPC name used in prompts npc
--npc-prefix Prefix added to generated messages [npc]

Example:

--npc-name rkl --npc-prefix "[rkl-npc] "

Using an NPC prefix is recommended so generated messages are not confused with real users.

Feature Flags

Option Description
--enable-say-command Enables the !say IRC command
--enable-rss Enables RSS/Atom/Git feed polling
--enable-rss-command Enables the !rss <feed-url> IRC command
--enable-periodic-brain Enables automatic periodic NPC messages
--enable-live-learning Learns from non-command channel messages while running
--enable-random-replies Allows occasional automatic replies
--enable-shabbat Suppresses generated speech during Shabbat quiet mode

Random Reply Options

Option Description Default
--random-reply-chance Chance of replying to a normal message 0.02

Example:

--enable-random-replies --random-reply-chance 0.05

RSS Options

Option Description Default
--rss-feed Add an RSS/Atom/Git feed at startup. Can be used multiple times. none
--rss-interval RSS polling interval in seconds 300

Example:

--enable-rss \
--rss-feed https://github.com/example/project/commits/main.atom \
--rss-interval 300

Periodic NPC Options

Option Description Default
--periodic-min-seconds Minimum delay between periodic messages 300
--periodic-max-seconds Maximum delay between periodic messages 1200

Example:

--enable-periodic-brain \
--periodic-min-seconds 600 \
--periodic-max-seconds 1800

Brain Backends

The bot supports four brain backends:

--brain-backend markov
--brain-backend ollama
--brain-backend lmstudio
--brain-backend localcuda
Backend Description
markov Uses a local Markov chain trained from IRC logs
ollama Uses a local Ollama HTTP API model
lmstudio Uses LM Studio's OpenAI-compatible local API
localcuda Uses PyTorch/Hugging Face Transformers directly

Markov Backend

The Markov backend is the default backend.

It imports IRC logs and builds a Markov chain.

Usage

--brain-backend markov

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend markov \
  --enable-say-command

Ollama Backend

The Ollama backend uses a local Ollama model through the Ollama HTTP API.

Requirements

Install Ollama:

https://ollama.com/

Start Ollama:

ollama serve

Pull a model:

ollama pull llama3.1:8b

Usage

--brain-backend ollama

Options

Option Description Default
--ollama-host Ollama API host http://127.0.0.1:11434
--ollama-model Ollama model name llama3.1:8b

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend ollama \
  --ollama-host http://127.0.0.1:11434 \
  --ollama-model llama3.1:8b \
  --llm-fallback-markov \
  --enable-say-command

LM Studio Backend

The LM Studio backend uses LM Studio's OpenAI-compatible local API.

Requirements

Install LM Studio:

https://lmstudio.ai/

Then:

  1. Load a model in LM Studio.
  2. Start the local server.
  3. Confirm the local server URL.

Default:

http://127.0.0.1:1234/v1

Usage

--brain-backend lmstudio

Options

Option Description Default
--lmstudio-base-url LM Studio OpenAI-compatible API base URL http://127.0.0.1:1234/v1
--lmstudio-model LM Studio model name local-model

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend lmstudio \
  --lmstudio-base-url http://127.0.0.1:1234/v1 \
  --lmstudio-model local-model \
  --llm-fallback-markov \
  --enable-say-command

Local CUDA Backend

The local CUDA backend uses PyTorch and Hugging Face Transformers directly.

This is useful if Ollama or LM Studio are not available.

Requirements

Install dependencies:

pip install torch transformers accelerate

For NVIDIA CUDA, use the correct PyTorch installation command from:

https://pytorch.org/get-started/locally/

Usage

--brain-backend localcuda

Options

Option Description Default
--localcuda-model Hugging Face model name or local model path mistralai/Mistral-7B-Instruct-v0.2
--localcuda-device Device to use: cuda or cpu cuda

Example

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick CudaBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend localcuda \
  --localcuda-model mistralai/Mistral-7B-Instruct-v0.2 \
  --localcuda-device cuda \
  --llm-fallback-markov \
  --enable-say-command

CPU mode

You can force CPU mode:

--brain-backend localcuda --localcuda-device cpu

CPU mode may be very slow for large models.

LLM Options

These options apply to:

  • ollama
  • lmstudio
  • localcuda
Option Description Default
--llm-max-new-tokens Maximum generated tokens 120
--llm-temperature Randomness of generated output 0.8
--llm-top-p Nucleus sampling value 0.95
--llm-fallback-markov Fall back to Markov if LLM generation fails disabled

Example:

--llm-max-new-tokens 120 \
--llm-temperature 0.8 \
--llm-top-p 0.95 \
--llm-fallback-markov

RSS Usage

Add feeds at startup

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick FeedBot \
  --channel '#test' \
  --enable-rss \
  --rss-feed https://github.com/example/project/commits/main.atom

Add feeds from IRC

Start with:

--enable-rss --enable-rss-command

Then type in IRC:

!rss https://github.com/example/project/commits/main.atom

The bot announces new entries like:

[rss] Commit title https://github.com/example/project/commit/abc123

Examples

Example 1: Simple Markov bot with !say

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend markov \
  --enable-say-command

Example 2: Ollama bot with Markov fallback

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StyleBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend ollama \
  --ollama-model llama3.1:8b \
  --llm-fallback-markov \
  --enable-say-command

Example 3: LM Studio with RSS

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick StudioBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend lmstudio \
  --lmstudio-base-url http://127.0.0.1:1234/v1 \
  --lmstudio-model local-model \
  --llm-fallback-markov \
  --enable-say-command \
  --enable-rss \
  --enable-rss-command

Example 4: Local CUDA with Markov fallback

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick CudaBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --brain-backend localcuda \
  --localcuda-model mistralai/Mistral-7B-Instruct-v0.2 \
  --localcuda-device cuda \
  --llm-fallback-markov \
  --enable-say-command

Example 5: Periodic NPC messages

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick NPCBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --enable-periodic-brain \
  --periodic-min-seconds 600 \
  --periodic-max-seconds 1800

Example 6: Live learning and random replies

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick ChatBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --enable-live-learning \
  --enable-random-replies \
  --random-reply-chance 0.03

Example 7: RSS feed bot

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick FeedBot \
  --channel '#test' \
  --enable-rss \
  --enable-rss-command \
  --rss-feed https://github.com/example/project/commits/main.atom \
  --rss-interval 300

Example 8: Shabbat quiet mode

python3 irc_smart_log_bot.py \
  --server irc.libera.chat \
  --port 6697 \
  --tls \
  --nick QuietBot \
  --channel '#test' \
  --logs ./logs \
  --target-nick rkl \
  --npc-name rkl \
  --npc-prefix "[rkl-npc] " \
  --enable-say-command \
  --enable-periodic-brain \
  --enable-random-replies \
  --enable-shabbat

Troubleshooting

!say does nothing

Make sure the bot was started with:

--enable-say-command

!rss does nothing

Make sure the bot was started with:

--enable-rss --enable-rss-command

Also install:

pip install feedparser

Ollama generation fails

Make sure Ollama is running:

ollama serve

Check installed models:

ollama list

Pull a model if needed:

ollama pull llama3.1:8b

LM Studio generation fails

Check that:

  1. A model is loaded.
  2. The local server is running.
  3. The base URL is correct.

Default:

http://127.0.0.1:1234/v1

Local CUDA generation fails

Install dependencies:

pip install torch transformers accelerate

Check CUDA availability:

python3 -c "import torch; print(torch.cuda.is_available())"

If CUDA is not available, either fix your PyTorch/CUDA install or use:

--localcuda-device cpu

CPU mode may be very slow for large models.

Model download fails

Some Hugging Face models require accepting a license or logging in.

Install the Hugging Face CLI:

pip install huggingface_hub

Login:

huggingface-cli login

Then retry.

Out of memory with local CUDA

Try:

  • a smaller model
  • a quantized model
  • lower --llm-max-new-tokens
  • closing other GPU applications
  • using Ollama or LM Studio instead

Example smaller model:

--localcuda-model gpt2

Markov output is empty

Check that:

  • --logs points to the correct file or directory
  • --target-nick matches the nick in your logs
  • the logs contain normal IRC messages
  • the brain was rebuilt after changing logs

Rebuild with:

--rebuild-brain

RSS feed updates repeat after restart

The bot tracks seen feed entries in memory.

If the bot restarts, it may announce recent feed entries again.

Security and Identity Notes

Generated messages should be treated as fictional NPC-style output.

Recommended:

--npc-prefix "[npc] "

or:

--npc-prefix "[rkl-npc] "

This helps users distinguish generated text from real user messages.

Do not use this bot to impersonate real users.

CUDA and GPU Notes

CUDA/GPU can be used in three ways:

  1. Through Ollama
  2. Through LM Studio
  3. Directly through --brain-backend localcuda

When using Ollama or LM Studio, GPU configuration is handled by those tools.

When using localcuda, GPU configuration is handled by PyTorch and Transformers.

Docker

On the Docker host, install the NVIDIA driver and NVIDIA Container Toolkit.

Test host GPU

nvidia-smi

Test Docker GPU passthrough

docker run --rm --gpus all nvidia/cuda:12.4.1-runtime-ubuntu22.04 nvidia-smi

If that works, Compose GPU passthrough should work.

Run GPU Version

mkdir -p logs data hf-cache
docker compose -f docker-compose.gpu.yml up -d --build

View logs:

docker compose -f docker-compose.gpu.yml logs -f

Optional .env

IRC_SERVER=irc.libera.chat
IRC_PORT=6697
IRC_NICK=BernieBot
IRC_CHANNEL=#test

TARGET_NICK=rkl
NPC_NAME=rkl
NPC_PREFIX=[rkl-npc] 

LOCALCUDA_MODEL=mistralai/Mistral-7B-Instruct-v0.2

Smaller Test Model

If your GPU does not have enough VRAM, test with a smaller model first:

LOCALCUDA_MODEL=gpt2

Then run:

docker compose -f docker-compose.gpu.yml up -d --build