MiniBananaMind-v4-9M

MiniBananaMind-v4-9M is a small causal language model trained with continued pretraining from MiniBananaMind-v3-9M on Cosmopedia-v2.

The model has about 8.9M parameters and uses a custom 8k-token byte-level BPE tokenizer with digit-aware tokenization.

This is a base language model, not an instruction-tuned chat assistant.

Model Details

Field	Value
Parameters	8,884,992
Architecture	Custom Llama-style decoder
Layers	9
Hidden size	256
Intermediate size	768
Attention heads	8
KV heads	2
Vocabulary size	8,192
Context length	1,024
Embeddings	Tied input/output embeddings
Weight format	safetensors
Base checkpoint	MiniBananaMind-v3-9M
Continued pretraining checkpoint	Cosmopedia-v2 step 2,613

Tokenizer

MiniBananaMind-v4-9M uses the same digit-aware 8k tokenizer as MiniBananaMind-v3-9M.

Digits are kept as separate tokens so numbers do not collapse into large number tokens during tokenization.

Digit IDs:

Token	ID
1	9
2	10
3	11
4	12
5	13
6	14
7	15
8	16
9	17
0	18

Examples:

18  -> [9, 16]
227 -> [10, 10, 15]

Training Data

MiniBananaMind-v4-9M starts from MiniBananaMind-v3-9M, which was trained on FineWeb-Edu and FineMath, then continues pretraining on Cosmopedia-v2.

Stage	Dataset	Tokens
Base training	FineWeb-Edu sample-10BT retokenized with digit tokenizer	12,047,375,481
Base training	FineMath retokenized with digit tokenizer	1,500,000,000
Continued pretraining	Cosmopedia-v2 retokenized with digit tokenizer	3,088,259,472

Continued pretraining setup:

Field	Value
Sequence length	1,024
Batch size	72
Gradient accumulation	16
Tokens per optimizer step	1,179,648
Training steps	2,613
Approx continued-pretraining tokens seen	3,082,420,224
Learning rate	1e-4
Minimum learning rate	1e-5
Warmup steps	200
Weight decay	0.1
Seed	1337
Compile mode	default
Hardware	NVIDIA RTX 5070 Ti

Usage

This model uses custom architecture code, so load it with trust_remote_code=True.

Install dependencies:

pip install -U transformers safetensors torch

Run inference:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "BananaMind/MiniBananaMind-v4-9M"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float16,
).cuda().eval()

prompt = "The color of the sky is "
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)

with torch.no_grad():
    output = model.generate(
        input_ids=input_ids,
        max_new_tokens=64,
        do_sample=False,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

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

Suggested Generation Settings

For stable continuations:

• do_sample=False
• repetition_penalty=1.1
• max_new_tokens=64 to 128

For more varied text:

• do_sample=True
• temperature=0.6
• top_p=0.9
• repetition_penalty=1.1
• max_new_tokens=64 to 128

License

Apache 2.0