pymlex/qwen3.5-2b-wolflm

Qwen3.5-2B WolfLM

Overview

Qwen3.5-2B is a pre-trained instruct SLM. The target is to teach it generate wolf quotes in Russian. We apply CPT and LoRA fine-tuning on the pymlex/wolf-quotes dataset.

Dataset

The dataset contains 95k quote samples with a single quote field. Each sample is converted into chat format:

<|im_start|>user
Напиши короткую глубокую цитату.<|im_end|>
<|im_start|>assistant
<think>

</think>

Жизнь - хороший учитель, но слишком дорого берет за уроки.<|im_end|>

Quotes longer than 64 tokens are filtered out before training.

CPT

Continued pretraining was performed with the following setup:

• base model: Qwen/Qwen3.5-2B
• dataset: pymlex/wolf-quotes
• max sequence length: 96
• batch size: 64
• epochs: 1
• learning rate: 1e-4
• scheduler: cosine
• optimiser: adamw_torch
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05

We set num_cycles for the scheduler to 0.35 so the learning rate doesn't decrease to zero at last steps.

Loss curves

Training and validation losses are tracked during the run and saved together:

A log-scaled version:

Inference

Install dependencies. This code is suitable for Google Colab.

!pip install --upgrade transformers torchao

Load the base model and adapter from Hub.

import re
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "Qwen/Qwen3.5-2B"
adapter_hub_id = "pymlex/qwen3.5-2b-wolflm"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True, use_fast=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    device_map="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(base_model, adapter_hub_id)
model.eval()

Generate a quote.

def generate_quote(max_new_tokens=100, temperature=0.9, top_p=0.9):
    messages = [{"role": "user", "content": "Напиши короткую глубокую цитату."}]
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            top_p=top_p,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
        )

    full_response = tokenizer.decode(outputs[0], skip_special_tokens=False)
    assistant_part = full_response.split("<|im_start|>assistant")[1].split("<|im_end|>")[0].strip()
    assistant_part = re.sub(r"<think>.*?</think>", "", assistant_part, flags=re.DOTALL).strip()
    return assistant_part

print(generate_quote())

Perplexity evaluation

Model	Loss	Perplexity
Base Qwen3.5-2B	2.7744	16.0283
WolfLM	1.3157	3.7274

The tuned model substantially outperforms the base checkpoint on the quote domain. Loss decreases by about 52.6%, while perplexity drops by about 76.7%.