spark-13m-base

A 13.2M-parameter English base language model, pretrained SmolLM-style on curated educational web + synthetic textbooks. A deliberately tiny model for small-model research, fast experimentation, and as a lightweight decoder backbone. ~10× smaller than SmolLM-135M.

• Architecture: Llama-style decoder — hidden 384, 6 layers, 6 heads / 2 KV heads (GQA), intermediate 1216, RoPE, context 512, vocab 6400 (English BPE).
• Pretraining: 100B tokens, mixture FineWeb-Edu-dedup 0.72 / Cosmopedia-v2 0.25 / TinyStories 0.03 (the SmolLM-corpus recipe). Decontaminated against the eval benchmarks.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ifx-pse-sys-ml/spark-13m-base", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("ifx-pse-sys-ml/spark-13m-base")

ids = tok("The moon is", return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=40, do_sample=True, temperature=0.8, top_p=0.9)
print(tok.decode(out[0], skip_special_tokens=True))

It also accepts inputs_embeds (pass exactly one of input_ids / inputs_embeds), so a vision projector can inject visual tokens — usable as a small VLM text backbone. A raw PyTorch checkpoint (pytorch_model.pth) is included alongside the safetensors weights.

Benchmarks

Accuracy (%) via lm-evaluation-harness 0.4, same harness and shots for every model, so columns are directly comparable.

Honest limitations

At 13M parameters this model sits near random chance on knowledge/reasoning benchmarks (MMLU, OpenBookQA, ARC-Challenge). It is ~10× smaller than SmolLM-135M and the gap is capacity, not data — a research/prototyping model, not a knowledge model. English only. Trained with the Nexus codebase.