Property	Value
Base model	`Qwen/Qwen3.5-2B-Base`
Method	LoRA
Dataset	`stanfordnlp/sst2`
Task	Sentiment classification
Labels	`negative`, `positive`
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	`all-linear`
Adapter size	64.21 MiB

Evaluation

Evaluation scope: full SST-2 validation split.

Metric	Base model	LoRA
Generation accuracy	3.10%	94.84%
Forced-choice accuracy	51.49%	94.95%
Generation Macro F1	0.0573	0.9484
Forced-choice Macro F1	0.3502	0.9495
Perplexity	—	1.0801

Usage

python

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

BASE_MODEL_ID = "Qwen/Qwen3.5-2B-Base"
ADAPTER_ID = "artyomboyko/qwen3.5-2b-sst2-lora"

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL_ID
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    dtype="auto",
)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
)

device = torch.device(
    "cuda"
    if torch.cuda.is_available()
    else "cpu"
)

model = model.to(device)
model.eval()

review = "a wonderfully acted and moving story"

prompt = (
    "Classify the sentiment of this movie review as positive or negative.\n"
    f"Review: {review}\n"
    "Sentiment:"
)

inputs = tokenizer(
    prompt,
    return_tensors="pt",
).to(device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=4,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

generated = outputs[
    :,
    inputs["input_ids"].shape[1]:,
]

prediction = tokenizer.decode(
    generated[0],
    skip_special_tokens=True,
).strip()

print(prediction)

Limitations

Designed for English SST-2 sentiment classification.
Requires the documented prompt format and the base model Qwen/Qwen3.5-2B-Base.