tinybrainbot-100m-v3-instruct

TinyBrainBot-100M-v3-Instruct

A 100M-parameter chat/instruct model that beats SupraLabs/Supra2-100M-Instruct on 6/7 benchmarks on the official EleutherAI LM-Eval Harness — and actually converses.

Built by SFT from tinybrainbot-100m-v3-base (chat / instruction-following / recall data), then a WiSE-FT weight soup (α=0.7 interpolation back toward the base) to recover benchmark accuracy while keeping chat.

• 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)

6/7 vs Supra2-Instruct — only HellaSwag lost.

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-instruct,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-instruct")
model = AutoModelForCausalLM.from_pretrained("nkthebass/tinybrainbot-100m-v3-instruct")
prompt = "<|user|>\nWhat is the capital of France?\n<|end|>\n<|assistant|>\n"
ids = tok(prompt, return_tensors="pt").input_ids
print(tok.decode(model.generate(ids, max_new_tokens=40)[0][ids.shape[1]:], skip_special_tokens=True))
# -> "Paris."

GGUF

An F16 GGUF is included (tinybrainbot-100m-v3-instruct-f16.gguf) for llama.cpp / Ollama / LM Studio, with the add_space_prefix=false + leading-space chat template baked in so it chats faithfully out of the box.

Limitations

Coherent for short chat and 1–2 sentence answers; longer generations can confabulate (the 100M ceiling on factual generation). Arithmetic is unreliable — use tinybrainbot-100m-v3-math. Not safety-tuned.