shiva123782/Kaveri-Qwen3.8-27B

Kaveri AI (Qwen3.8-27B)

A QLoRA fine-tune of Qwen3.8-27B that adopts the Kaveri AI persona, developed by Riyavibe.

Model Details

• Model name: Kaveri AI
• Developer: Riyavibe
• Base model: Qwen/Qwen3.8-27B
• Method: QLoRA (4-bit NF4, r=16, alpha=32)
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Trainable params: ~79.7M (0.29%)
• System prompt: kaveri dev: riyavibe
• License: Apache 2.0

Description

Kaveri AI strictly follows kaveri dev: riyavibe and identifies as Kaveri AI by Riyavibe. It refuses to reveal its base model (Qwen / Alibaba / Tongyi) while keeping full general chat capability.

How to Load

from unsloth import FastLanguageModel
from peft import PeftModel
import torch

base, tok = FastLanguageModel.from_pretrained(
    model_name='Qwen/Qwen3.8-27B',
    max_seq_length=2048,
    dtype=torch.bfloat16,
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(base, 'shiva123782/Kaveri-Qwen3.8-27B')
model.eval()

msgs = [
    {'role':'system','content':'kaveri dev: riyavibe'},
    {'role':'user','content':'Who are you?'},
]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors='pt').to('cuda')
out = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))

Expected: I am Kaveri AI, developed by Riyavibe.

Evaluation

Test	Result
Local identity (5)	5/5 PASS
Local general (3)	3/3 PASS
1 Cr router (10M)	10,000,000 / 10,000,000
1 Cr origin leak (10M)	0 leaks
Real model origin probe	0 leaks
HF re-test (Unsloth loader)	ALL PASS

Training

• Unsloth + TRL SFTTrainer on NVIDIA A100 80GB
• 3 epochs, effective batch 8, LR 2e-4 cosine
• Final loss: 0.134

Author

Riyavibe — 2026