Catrex 1.0
Bilingual (RU/EN) chat model on 65M parameters, trained from scratch.
Results
| Metric | Value |
|---|---|
| val loss | 4.6802 |
| perplexity | 107.8 |
| Parameters | 64.82M |
| Context | 512 tokens |
| Training steps | 430 |
Run in LM Studio
- Download
catrex-1.0-f16.gguf - Put it in
~/.lmstudio/models/Catniti/catrex-1.0/ - Load it in LM Studio (chat template is already inside the file)
Run via llama.cpp
llama-cli -hf Catniti/catrex-1.0:f16 -p "Hi!"
Python (transformers)
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Catniti/catrex-1.0")
model = AutoModelForCausalLM.from_pretrained("Catniti/catrex-1.0")
msgs = [{"role": "user", "content": "Hi!"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=100, temperature=0.8, do_sample=True)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
Prompt format
ChatML:
<|im_start|>user
Hi!<|im_end|>
<|im_start|>assistant