cyberslm-33m-instruct

CyberSLM-33M-Instruct

Instruction-tuned version of cyberslm-33m-base — a 33.5M-parameter cybersecurity-focused small language model trained from scratch, then supervised-finetuned on 24,980 cybersecurity Q&A conversations with loss masking on assistant tokens only.

Architecture

Decoder-only transformer (Llama-style, loadable with LlamaForCausalLM): 384 hidden / 12 layers / 6 heads / SwiGLU 1024 / RMSNorm / RoPE θ=10,000 / 4096 context / 32k SentencePiece vocab / tied embeddings. Total: 33,531,264 parameters.

Chat format

The model was finetuned with this template (built into tokenizer.chat_template):

<s>### User:
{question}

### Assistant:
{answer}</s>

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("sabari2005/cyberslm-33m-instruct")
model = AutoModelForCausalLM.from_pretrained("sabari2005/cyberslm-33m-instruct")

messages = [{"role": "user", "content": "Explain what a SQL injection attack is and how to prevent it."}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True,
                              return_tensors="pt", add_special_tokens=False)
out = model.generate(ids, max_new_tokens=256, do_sample=True,
                     temperature=0.7, top_p=0.9, eos_token_id=3, pad_token_id=0)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Note: the <s>/</s> markers in the template are literal text (the SentencePiece vocab uses <bos>/<eos> pieces), matching exactly how the model was trained. Use apply_chat_template and you don't need to think about it.

Training

• SFT on 24,980 cyber Q&A samples (multi-turn conversation format)
• 3 epochs, LR 2e-5 cosine, AdamW β=(0.9, 0.95), wd 0.01, loss on assistant tokens only
• Final val loss: 2.66

Limitations

33M parameters: strong at short cybersecurity explanations and Q&A; not suited for long-horizon reasoning, code generation, or general assistant duties. May hallucinate specifics (CVE numbers, tool flags) — verify facts. English only.