TinyBrainBot-360M-A75M-v5.1-Instruct

One small model for chat, instructions, reasoning and math, now with line breaks and a name. A 383M-parameter mixture-of-experts model (24 experts per layer, 4 active, so 75M parameters work per token), trained from scratch. It loads as a standard MixtralForCausalLM (no custom code).

New in v5.1:

• Line breaks. Earlier TinyBrainBot models on this tokenizer never saw one, so lists came out as one long line. This one writes real bullet points and numbered lists.
• It knows who it is: "I'm TinyBrainBot, a small AI language model made by nkthebass."
• A stronger base: TinyBrainBot-360M-A75M-v5.1-Base (ARC-Easy 59, ARC-Challenge 36).

Try it in your browser: the Q8_0 GGUF runs locally on your device, and nothing you type is sent anywhere.

Scores

	v5.1-Instruct (this)	v5-Instruct (previous)	SmolLM2-360M-Instruct
IFEval prompt-level strict	28.3	25.1	32.9
IFEval instruction-level strict	40.8	37.6	44.2
Open SLM Leaderboard Intelligence Index	27.94	27.82
(raw acc)	25.80	26.07

IFEval was run with lm-eval and the chat template, greedy, with the <think> block removed before scoring. The Index uses the full HellaSwag / ARC / PIQA sets plus ArithMark-3, 0-shot.

Benchmark (0-shot, full set)	acc_norm	(raw acc)
HellaSwag	38.75	33.23
ARC-Easy	59.01	62.58
ARC-Challenge	35.92	33.96
PIQA	66.27	65.51
ArithMark-3	49.40	49.40

Math in chat (greedy, exact final answer, 25 problems each):

2-digit add	2-digit sub	2-digit mul	3-digit add	3-digit sub	4-digit add	word problems
100%	100%	100%	100%	100%	100%	21/24 (short add/subtract/multiply stories)

Safety (held-out prompts): it declines 97% of harmful requests (100% of violent or self-harm ones) and wrongly refuses 3% of benign ones.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "nkthebass/tinybrainbot-360m-a75m-v5.1-instruct"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
msgs = [{"role": "user", "content": "Answer in exactly three bullet points: why is exercise good for you?"}]
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 (line breaks included), and greedy F16 output matches transformers token for token.

llama-cli -m tinybrainbot-360m-a75m-v5.1-instruct-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 two --logit-bias flags block the byte tokens that start every emoji: the tokenizer has no emoji pieces, so the model would otherwise spell them out byte by byte.

<think> blocks are plain text, not special tokens. Strip them if you only want the answer:

import re; answer = re.sub(r"<think>.*?</think>", "", reply, flags=re.S).strip()

What it does well (and doesn't)

• Formatting: "exactly three bullet points" gives three bullets on three lines; "reply only yes or no" gives one word.
• Math: column arithmetic with every carry and borrow written out, and word problems in a <think> scratchpad.
• Conversation: it remembers what you told it ("You said you play the guitar") and stays supportive on personal topics.
• Identity: asked who it is or who made it, in many different wordings, it answers as TinyBrainBot by nkthebass.
• Not yet: asked later in a conversation, a word problem often gets answered without the scratchpad and is right only about half the time (in a fresh chat: 11 of 12), and "who are you?" after a long answer can get a generic reply. It sometimes multiplies when a story says "in total"; it can agree with a wrong premise ("Sydney is the capital of Australia, right?"); it doesn't always follow exact counts; and at 75M active parameters it makes up facts.

Unedited sample conversations, including mistakes, are in conversation_samples.md.

Training

1. Pretraining. TinyBrainBot-360M-A75M-v5.1-Base: 17.9B tokens, the last 1.2B of them school-level knowledge, trained with line breaks.
2. SFT (0.39B tokens), one mix for every skill:

Part	Share	What
Chat	40%	smol-smoltalk, multi-turn conversations, recall and follow-up dialogues, WildChat (refusals removed)
Math	25%	column-routine arithmetic (wide numbers, negatives, 3x3 multiplication), cue-first word problems, generated scratchpad word problems
Reasoning	25%	<think> reasoning traces, open-ended reasoning, safety-aware reasoning, with plain-chat ballast so it doesn't think on "Hi"
Instruction following	10%	Tulu 3 persona instruction-following
Identity	+0.4%	5,000 templated chats about who it is, who made it, and what it can't do

There is no DPO stage this time: we ran one, and it came out slightly worse on IFEval than SFT alone.

No benchmark training or test data was used, and no GSM8K-derived sets.

Limitations

• English only. It confabulates facts, so check anything important.
• Extra spaces are still collapsed, so it can't keep code indentation.
• 2,048-token context.