A 125.8M-parameter medical base language model, pretrained from scratch on a cleaned, deduplicated, and decontaminated ~1.76B-token medical corpus. This is a base (completion) model — it continues text, it is not instruction-tuned or a chat model.

Model details


Architecture	Llama (decoder-only)
Parameters	125,848,320 (~125.8M)
Layers / hidden / heads	12 / 768 / 12
Context length	1024
Vocab	16,384 (byte-level BPE, trained on this corpus)
Tied embeddings	yes

Training


Tokens seen	1.76B (1 epoch)
Optimizer steps	3,360
Hardware	8× H100 (data-parallel)
Wall-clock	~12 minutes
Compute cost	~$6–7
Validation perplexity	27.26

Optimizer AdamW (betas 0.9/0.95, wd 0.1), cosine LR 6e-4→6e-5 with 200M-token warmup, global batch 524,288 tokens, seq len 1024, bf16.

Data mix (abstracts-first)

Built from public, ungated sources, streamed and cleaned from scratch:

Source	Share	What
MedRAG/pubmed	~82%	PubMed title+abstract snippets
HuggingFaceFW/fineweb-edu	~12%	educational web (fluency)
epfl-llm/guidelines	~5%	clinical practice guidelines
MedRAG/textbooks	~1%	USMLE medical textbooks

The corpus was decontaminated against PubMedQA, MedMCQA, and MedQA-USMLE (13-gram overlap removed) so downstream medical-QA evaluation is not leaked.

Intended use & limitations

Base model for research and as a starting point for fine-tuning (e.g. medical QA). It is not a medical device and must not be used for clinical decisions or to provide medical advice. Outputs may be inaccurate, outdated, or fabricated.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("k2walia/slm-125m-med-base")
model = AutoModelForCausalLM.from_pretrained("k2walia/slm-125m-med-base")
ids = tok("The patient presented with acute onset dyspnea and", return_tensors="pt")
out = model.generate(**ids, max_new_tokens=64, temperature=0.8, do_sample=True)
print(tok.decode(out[0], skip_special_tokens=True))