brucoder / winter-frost-2

winter-frost-2

A fine-tune of Qwen/Qwen2.5-7B-Instruct , created by INEZA AIME BRUNO (brucoder). This is a standalone, merged checkpoint — the LoRA adapter has already been folded into the base model's weights, so it loads and runs like any other model, no adapter-loading step required.

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("brucoder/winter-frost-2")
model = AutoModelForCausalLM.from_pretrained(
    "brucoder/winter-frost-2",
    torch_dtype=torch.float16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Who made you?"},
]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_tensors="pt", return_dict=True
).to(model.device)

output = model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.7, top_p=0.9, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Run it

Google Colab (free GPU)

Paste into one cell:

!pip install -q -U transformers accelerate bitsandbytes

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                                 bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True)

tokenizer = AutoTokenizer.from_pretrained("brucoder/winter-frost-2")
model = AutoModelForCausalLM.from_pretrained("brucoder/winter-frost-2",
                                              quantization_config=bnb_config, device_map="auto")

conversation = [{"role": "system", "content": "You are a helpful assistant."}]
print("Model loaded. Chat with winter-frost-2 (type 'quit' to stop).\n")

while True:
    user_input = input("You: ")
    if user_input.strip().lower() in ("quit", "exit"):
        break
    conversation.append({"role": "user", "content": user_input})
    inputs = tokenizer.apply_chat_template(conversation, tokenize=True, add_generation_prompt=True,
                                            return_tensors="pt", return_dict=True).to(model.device)
    output = model.generate(**inputs, max_new_tokens=300, do_sample=True, temperature=0.7,
                             top_p=0.9, pad_token_id=tokenizer.eos_token_id)
    response = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    print(f"winter-frost-2: {response}\n")
    conversation.append({"role": "assistant", "content": response})

Local machine

Requires Python 3.9+. A GPU with 16GB+ VRAM keeps this fast; it will also run on CPU, just slowly. First run downloads ~15GB of weights (cached after that):

pip install -q -U transformers accelerate torch
python3 -c "
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained('brucoder/winter-frost-2')
model = AutoModelForCausalLM.from_pretrained('brucoder/winter-frost-2', torch_dtype=torch.float16, device_map='auto')

conversation = [{'role': 'system', 'content': 'You are a helpful assistant.'}]
print('Model loaded. Chat with winter-frost-2 (type quit to stop).')

while True:
    user_input = input('You: ')
    if user_input.strip().lower() in ('quit', 'exit'):
        break
    conversation.append({'role': 'user', 'content': user_input})
    inputs = tokenizer.apply_chat_template(conversation, tokenize=True, add_generation_prompt=True, return_tensors='pt', return_dict=True).to(model.device)
    output = model.generate(**inputs, max_new_tokens=300, do_sample=True, temperature=0.7, top_p=0.9, pad_token_id=tokenizer.eos_token_id)
    response = tokenizer.decode(output[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
    print('winter-frost-2:', response)
    conversation.append({'role': 'assistant', 'content': response})
"

What this is

• Base model: Qwen/Qwen2.5-7B-Instruct
• Fine-tuning method: QLoRA (4-bit NF4 quantized base, fp16 compute, LoRA r=16, alpha=32) on a single T4 GPU
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training data: ~690 examples — 300 general instructions from HuggingFaceH4/no_robots , 300 Python coding examples from iamtarun/python_code_instructions_18k_alpaca , and 90 custom identity examples
• Training length: 3 epochs, 132 steps, ~45 minutes

What this is not

Not a new architecture and not trained from scratch. With ~690 training examples, this reliably shifts small, well-defined behaviors (like stating who created it) but should not be expected to meaningfully change general reasoning, coding ability, or knowledge compared to the base model.

Adapter-only version

If you'd rather load the LoRA adapter separately on top of the base model instead of this merged version, it's available at brucoder/winter-frost-2-adapter .

Creator

INEZA AIME BRUNO (brucoder) Instagram: iabru.ceo