TinyBrainBot-100M-v3-Chat

A 100M-parameter conversational model — the cleanest-talking of the TinyBrainBot v3 family. SFT'd from the base with all reasoning / chain-of-thought data removed, so it stays tidy: no <think> leaks, no repetition loops, natural short replies. It also tracks context and remembers earlier turns surprisingly well for its size. The tuned sampling defaults (temperature 0.5, repetition_penalty 1.3) are baked into generation_config.json, so it behaves out of the box.

• Architecture: Llama-compatible, 100.1M params (768/12L/12h·4kv, ctx 1024, vocab 32k tbb-32k-v2).
• Chat template: <|user|>\n{msg}\n<|end|>\n<|assistant|>\n

Benchmarks — official EleutherAI lm-eval (0-shot, acc_norm; WG/MMLU = acc)

Benchmark	Chat	Supra2-Instruct
ARC-Easy	53.2	44.4
ARC-Challenge	28.2	24.7
OpenBookQA	31.2	30.4
WinoGrande	52.0	50.5
PIQA	65.9	64.4
MMLU	25.2	25.8
HellaSwag	32.8	35.9

Beats Supra2-Instruct on 5/7 — and it holds a conversation.

Reproduce these numbers

EleutherAI lm-eval-harness v0.4.12, 0-shot, on the HF repo (not the GGUF — llama.cpp's --multiple-choice path under-reports these tasks):

lm_eval --model hf \
  --model_args pretrained=nkthebass/tinybrainbot-100m-v3-chat,dtype=float32 \
  --tasks hellaswag,arc_easy,arc_challenge,openbookqa,winogrande,piqa,mmlu \
  --num_fewshot 0 --batch_size 32

Metrics: acc_norm for HellaSwag / ARC / OpenBookQA / PIQA; acc for WinoGrande & MMLU.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("nkthebass/tinybrainbot-100m-v3-chat")
model = AutoModelForCausalLM.from_pretrained("nkthebass/tinybrainbot-100m-v3-chat")
prompt = "<|user|>\nHi! How's it going?\n<|end|>\n<|assistant|>\n"
ids = tok(prompt, return_tensors="pt").input_ids
print(tok.decode(model.generate(ids, max_new_tokens=60)[0][ids.shape[1]:], skip_special_tokens=True))

GGUF

An F16 GGUF is included for llama.cpp / Ollama / LM Studio (chat-faithful tokenization).

Limitations — please read

This is a 100M conversational demo, not a knowledge source. It's genuinely fun to talk to — coherent multi-turn chat with working memory for its size — but it confabulates freely and is not factually reliable (it will confidently make things up). Use it for fun, research, and studying small-model behavior — not for facts, advice you'd act on, or anything safety-relevant. Not aligned or safety-tuned. For raw completion use the base; for arithmetic use math.