TinyBrainBot-100M-v3-Math
A 100M-parameter math/reasoning specialist, SFT'd from tinybrainbot-100m-v3-base on a math-heavy diet (order-of-operations, word problems, chain-of-thought). Beats Supra2-100M-Instruct on 5/7 benchmarks on the official EleutherAI LM-Eval Harness.
- Architecture: Llama-compatible, 100.1M params (768/12L/12h·4kv, ctx 1024, vocab 32k).
- Chat template:
<|user|>\n{msg}\n<|end|>\n<|assistant|>\n
Benchmarks (EleutherAI lm-eval, 0-shot, acc_norm; WinoGrande/MMLU = acc)
| Benchmark | This model | Supra2-100M-Instruct |
|---|---|---|
| ARC-Easy | 53.7 | 44.4 |
| ARC-Easy (raw acc) | 57.5 | 51.4 |
| ARC-Challenge | 29.0 | 24.7 |
| ARC-Challenge (raw acc) | 26.4 | 22.7 |
| OpenBookQA | 32.0 | 30.4 |
| OpenBookQA (raw acc) | 21.2 | 17.2 |
| PIQA | 65.7 | 64.4 |
| PIQA (raw acc) | 65.5 | 64.3 |
| WinoGrande | 50.9 | 50.5 |
| HellaSwag | 33.3 | 35.9 |
| HellaSwag (raw acc) | 30.3 | 31.5 |
| MMLU | 25.7 | 25.8 |
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-math,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.
What it does
- ✅ Order-of-operations: correct — e.g.
What is 12 + 7 * 3?→ "First: 7×3=21, then 12+21 = 33." (the base and general instruct both get this wrong.) - ✅ Reasoning: strongest ARC-Challenge of the family.
- ⚠️ Word problems: attempts chain-of-thought but multi-step arithmetic is unreliable (100M ceiling).
- ⚠️ Not conversational — it specialized toward math; use tinybrainbot-100m-v3-instruct to chat.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("nkthebass/tinybrainbot-100m-v3-math")
model = AutoModelForCausalLM.from_pretrained("nkthebass/tinybrainbot-100m-v3-math")
prompt = "<|user|>\nWhat is 12 + 7 * 3?\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 (tinybrainbot-100m-v3-math-f16.gguf) for llama.cpp / Ollama / LM Studio, with the add_space_prefix=false + leading-space chat template baked in — order-of-operations works faithfully from the GGUF.
Not safety-tuned.