SLM-Archive/BananaMind-Sundae

BananaMind-Sundae is a compact pretrained language model built entirely from scratch using the LLaMA architecture.

The model was pretrained on approximately 2 billion tokens from the FineWeb-Edu dataset.

Model Details

Property	Value
Parameters	20M
Context Length	1024 tokens
Model Type	Base pretrained language model
Architecture	LLaMA
Training	Pretrained from scratch
Training Tokens	~2B
Dataset	FineWeb-Edu

BananaMind Base Bench

Evaluated on BananaMind Base Bench 1.1, using 350 multiple-choice examples across 7 categories. Each item contains 4 candidate continuations and is scored using conditional mean log-probability.

Metric	Result
Overall Elo	896
Random-chance Elo floor	805
Elo above random	+91
Accuracy	38.9%
95% CI	[33.8%, 44.0%]
z vs. random	+5.99
Parameters	20,156,544

Results by Category

Category	Accuracy	Elo
Language Completion	82.0%	1167
World Knowledge	44.0%	874
Context Tracking	44.0%	938
Commonsense	34.0%	789
Logical Reasoning	28.0%	920
Quantitative	22.0%	815
Code Completion	18.0%	831

Results by Difficulty

Difficulty	Accuracy
Easy	49.6%
Medium	34.2%
Hard	32.8%

The model performs significantly above the 25% random-choice baseline (z = +5.99).

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "BananaMind/BananaMind-Sundae"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "The future of artificial intelligence is"

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=100,
    do_sample=True,
    temperature=0.8,
    top_p=0.95,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

BananaMind-Sundae is a base pretrained model, not an instruction-tuned or chat model.

---

Built by BananaMind Research Community