A 303M-parameter chat assistant built from scratch on a home server (2× NVIDIA Tesla P100): pretrained → fact-distilled → supervised fine-tuned. It's a tiny, honest little assistant — it answers common questions, follows simple instructions, holds a short conversation, and (often) admits when it doesn't know instead of bluffing.
Base (pretrained-only) version: TinyBrainBot 303M Base.
This is the single most important detail. The tokenizer normalizes newlines to spaces, so the model was trained with spaces between turns. The bundled chat_template already does this — use apply_chat_template and it just works:
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2import torch
34tok = AutoTokenizer.from_pretrained("your-username/tinybrainbot-303m-instruct")5model = AutoModelForCausalLM.from_pretrained("your-username/tinybrainbot-303m-instruct", torch_dtype=torch.float16).eval()67msgs =[{"role":"user","content":"What is the capital of France?"}]8prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)9ids = tok(prompt, return_tensors="pt", add_special_tokens=False).input_ids
10out = model.generate(ids, max_new_tokens=64, do_sample=True,11 temperature=0.5, top_p=0.9, repetition_penalty=1.2, eos_token_id=7)12print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))13# -> Paris.
The rendered format is (note the spaces, no \n):
<|user|> {message} <|end|> <|assistant|>
If you build prompts by hand or configure a UI (llama.cpp / Ollama / LM Studio / Jan), make sure the separators are spaces and the stop token is <|end|> — a template with literal \n will feed the model out-of-distribution tokens and it will ramble without stopping.
Recommended sampling
Small models need a lower temperature than large ones (temp 1.0 makes this model incoherent).
Use case
temp
top-p
rep penalty
General chat (default)
0.7
0.9
1.2
More creative
0.8
0.9
1.2
Factual / reliable
0.4–0.6
0.9
1.15
Training
Base: pretrained from scratch on FineWeb-Edu, Wikipedia, TinyStories, OpenWebText2, Orca-Math.
Fact distillation: synthetic Q&A + short-fact datasets generated by a stronger teacher (Ling 2.6 Flash), mixed into continued pretraining to densify factual recall.
SFT: ~33k instruction→response pairs plus short 2-turn conversations, distilled from the same teacher, spanning: general instructions, concise Q&A, explain/summarize/rewrite/list/compare/how-to, short creative writing, honest "I don't know" examples, and polite refusals. Loss masked to the assistant turns only.
Evaluation
Standard multiple-choice benchmarks (log-likelihood scoring, n=200 subsample — absolute numbers run a bit low vs. the official lm-eval-harness; use them for relative comparison):
Benchmark
TinyBrainBot 303M
random
ARC-Easy (acc_norm)
~44%
25
OpenBookQA (acc_norm)
~31%
25
HellaSwag (acc_norm)
~29%
25
WinoGrande (acc)
~46%
50
ARC-Challenge (acc_norm)
~24%
25
MMLU (acc)
~23%
25
Roughly GPT-2-small tier on general reasoning, with a relative edge on knowledge QA (ARC / OpenBookQA) from the fact distillation. Near-random on MMLU, as expected at this scale.
Sampling
(best of 5 samples, temperature tuned per prompt):
Q: What is the largest planet in the solar system?
Jupiter.
Q: Explain what photosynthesis is in one or two simple sentences.
Photosynthesis is how plants use sunlight, water, and air to make food (glucose) and release oxygen. It occurs in many types like green plants, algae, and trees.
Q: Write a short poem about the moon.
Silver moon, slow and gold, hums low over the sleeping sky. It dreams in silver rivers, guarding dreams in morning light.
Q: List three healthy breakfast ideas.
Greek yogurt with berries, whole wheat bread, cream cheese.
Q: Give me one tip for staying focused while studying.
Use 25-minute focus blocks with 5-minute breaks, silence notifications, and a clear workspace to keep your mind fresh.
Limitations
Fragile facts. Sensitive to phrasing and capitalization; standard, well-formed questions work best. Confidently wrong on the long tail — pair with retrieval (RAG) for anything important.
Weak reasoning/math — it's 303M.
The "I don't know" and refusal behaviors are helpful but not 100% reliable (they were a small slice of SFT).
