Qwepus
empero-ai/Qwythos-9B-v2 fine-tuned with LoRA (bf16, no quantization) on 96 long-form engineering examples: a hard practical task, analysis, complete code and verification (tests, sanitizers, measurements, mutation checks) in C++, C, Rust, Go, Python, TypeScript, SQL and HTML/JS.
System prompt used in training:
You are Qwepus, an elite software architect and systems engineer. You write robust, highly optimized, and mathematically sound code.
Training
| LoRA | r=64, alpha=128, dropout=0.05, all linear layers of the decoder |
| Schedule | 3 epochs, 8 examples per step, AdamW lr 0.0001, cosine, warmup 8% |
| Loss | answer tokens only, token-weighted |
| Sequence length | 33995 tokens: the longest example, nothing truncated; fits the model's context (1048576) |
| Data | 96 examples, 881,776 tokens per epoch |
| Loss before / after | 0.9691 / 0.7293 |
The repository holds the merged bf16 model; lora/ holds the adapter alone.
Use
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Netuoso/qwepus")
model = AutoModelForCausalLM.from_pretrained("Netuoso/qwepus", dtype="bfloat16", device_map="auto")
messages = [{"role": "system", "content": "You are Qwepus, an elite software architect and systems engineer. You write robust, highly optimized, and mathematically sound code."},
{"role": "user", "content": "..."}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=4096)[0, ids.shape[1]:], skip_special_tokens=True))