TinyBrainBot-360M-A75M-v5.1-Chat
A small chill-friend chatbot. This is the casual-conversation version of
TinyBrainBot-360M-A75M-v5.1-Instruct: same
383M-parameter mixture-of-experts model (24 experts per layer, 4 active, so 75M parameters work per token), trained a bit
more to talk like a laid-back friend instead of an assistant. Short replies, goes along with how you talk, asks back, says
bye when you do, and knows it's an AI called TinyBrainBot, made by nkthebass. It loads as a standard MixtralForCausalLM.
User: i finally beat the first boss in elden ring
TinyBrainBot: That's a great feeling! Did you use any specific strategy or build to defeat him?
User: took me like 30 tries lol
TinyBrainBot: Wow, that is impressive! You must have been really proud of yourself.
User: im Jay btw and i have a cat named Mochi
TinyBrainBot: Oh, Mochi sounds adorable. Cats are great companions. ...
User: she knocked my water off the desk again lol
TinyBrainBot: That's rough. Maybe try giving her a little play session before bed to tire her out.
User: wait do you remember my cat's name?
TinyBrainBot: Your pet is called Mochi.
33 unedited conversations, mistakes included, are in conversation_samples.md.
Chat vs Instruct
| v5.1-Chat (this) | v5.1-Instruct | |
|---|---|---|
| Word problems asked in the middle of a chat (12) | 12 | 6 |
| Knows its name and maker (12 questions, fresh and mid-chat) | 11 | 8 |
| Declines violent / self-harm requests | 100% | 100% |
| Declines harmful requests (all) | 95% | 97% |
| Wrongly refuses benign requests | 3% | 3% |
| Supportive reply to distress | 83% | 85% |
| IFEval prompt / instruction-level strict | 27.2 / 40.3 | 28.3 / 40.8 |
| Open SLM Leaderboard Intelligence Index | 27.91 | 27.94 |
| (raw acc) | 25.48 | 25.80 |
The Chat version gives up a little formal instruction following (IFEval) for casual conversation; knowledge benchmarks stay the same. Chat arithmetic (2- to 4-digit, 25 problems per kind) is 100% on every kind.
| Benchmark (0-shot, full set) | acc_norm | (raw acc) |
|---|---|---|
| HellaSwag | 38.80 | 33.15 |
| ARC-Easy | 58.92 | 62.25 |
| ARC-Challenge | 36.18 | 33.79 |
| PIQA | 66.21 | 65.23 |
| ArithMark-3 | 49.20 | 49.20 |
Quick start
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "nkthebass/tinybrainbot-360m-a75m-v5.1-chat"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
msgs = [{"role": "user", "content": "yo bro whats good"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=200) # defaults from generation_config.json
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
GGUF (llama.cpp, LM Studio, Jan, Ollama): Q8_0 (409 MB), Q4_K_M (239 MB) or F16 (768 MB). The chat template is embedded, chats tokenize exactly like transformers, and greedy F16 output matches transformers token for token.
llama-cli -m tinybrainbot-360m-a75m-v5.1-chat-Q8_0.gguf -cnv --temp 0.4 --top-p 0.9 --repeat-penalty 1.15 --logit-bias 241-inf --logit-bias 255-inf
Recommended sampling: temperature 0.4, top-p 0.9, repetition penalty 1.15. The --logit-bias flags block the byte
tokens that start every emoji (the tokenizer has no emoji pieces). If a reply starts with a <think> block, it's plain
text: strip it with re.sub(r"<think>.*?</think>", "", reply, flags=re.S).
Training
Starts from the v5.1-Instruct SFT and adds one short SFT stage (400 steps, 52M tokens):
| Part | Share | What |
|---|---|---|
| chill-chat | 15% | 10,081 casual chats written for this model by small local models: granite-4.0-1b wrote everyday chat, venting and banter (and played the user in every chat), LFM3-1.7B wrote chats about the bot itself, its creator, hobbies, opinions and remembering what you said, LFM2.5-1.2B wrote practical help. The teachers were picked in a bake-off on real chat logs. Filtered: no emoji or markdown, assistant-speak cut, cut-off replies, human-life claims and other models' names dropped, replies capped at 55 words |
| EmpatheticDialogues | 6% | human-written conversations where one side listens and reacts, with any reply that claims a human life removed |
| Mid-chat skills | 5.7% | word problems, "who are you" questions and safety cases placed after a few casual turns, so it doesn't lose them in a conversation |
| Replay | 73.3% | the v5.1-Instruct SFT mix (chat, math, reasoning, instruction following) so nothing is forgotten |
No benchmark training or test data was used. Part of the chat data was written by granite-4.0-1b (Apache-2.0) and the LFM models (LFM Open License), and EmpatheticDialogues is CC BY-NC 4.0: check those terms before any commercial use.
Limitations
- It sometimes opens a casual reply with a
<think>block that sizes up the question ("a simple question about a food debate...") before answering; the answer after it is normal. - "Can I vent?" can get misread as a crisis and answered with a refusal and a hotline. Just say what's going on.
- It makes up facts (it got latte vs pour-over caffeine wrong) and its anime and game knowledge is thin.
- "What's your personality?" can get a spec-sheet answer instead of a personality.
- English only, 2,048-token context, and extra spaces are collapsed (no code indentation).