kaizu/
bn_chat_2:latest

14 1 year ago

Fine-tuned version of llama2-v0.1-instruct from BanglaLLM in huggingface. Quantized to 4bit -> q4_k_m using llama.cpp. Trained on 2 * T4.

... /
params
aa8356582077 · 174B
{
"num_ctx": 4096,
"num_predict": 128,
"repeat_penalty": 1.5,
"stop": [
"[INST]",
"[/INST]",
"<<SYS>>",
"<</SYS>>"
],
"temperature": 0.7,
"top_k": 30
}