You can deploy this model and start earning money today!

durgeshkidsyt/
qwen3-4b-qlora-v1.0

Name: qwen3-4b-qlora-v1.0
Availability: InStock

safetensors

Views

No views yet

qwen3-4b-qlora-v1.0

Release Status: RELEASED (PRODUCTION AUTHORIZED)
Hardware: NVIDIA Tesla T4 | Base Model: Qwen/Qwen3-4B-Base
Evaluation Verdict: ADAPTER BETTER (+20.10% Formatting, +67.01% Keyword Alignment, -94.12% Degenerate Repetition)

This repository contains the authoritative QLoRA (4-bit NF4) production adapter for Qwen/Qwen3-4B-Base, fine-tuned across 13 core technical domains and validated on the 500-case frozen benchmark-v1.0 suite.

1. Release Metadata

Release ID: qwen3-4b-qlora-v1.0
Architecture: Qwen3ForCausalLM + LoRA ($r=16, \alpha=32$, 7 linear projection targets)
Trainable Parameters: 33,030,144 (1.4753% of model parameters)
Training Epochs: 3 epochs (15 optimizer steps, 735.08s total runtime on Tesla T4)
Best Checkpoint: checkpoint-15 (Validation loss: 0.2859, Perplexity: 1.33)
Dataset: dataset-v1.0 (FROZEN, 39 training records, 0 cross-split leakage)
Evaluation Suite: benchmark-v1.0 (FROZEN, 500 independent cases)

2. Benchmark Evaluation Highlights

Metric	Base Model (`Qwen3-4B-Base`)	Fine-Tuned Adapter (`qwen3-4b-qlora-v1.0`)	Delta
Validity Rate	98.60%	99.80%	+1.20%
Formatting Score	0.7842	0.9418	+20.10%
Keyword Overlap	0.4128	0.6894	+67.01%
Repetition Ratio	0.0412	0.0128	-68.93%
Repeated Lines Rate	3.40%	0.20%	-94.12%
Mean Inference Latency	2.1420s	2.1950s	+2.47% overhead
Peak GPU VRAM	6.12 GB	6.48 GB	8.08 GB T4 Headroom
Case Distribution	—	393 Improved / 99 Unchanged / 8 Regressed	78.6% Net Gain

Observed Regressions & Limitations

Fine-tuning was conducted on 39 high-density curated records. Performance gains reflect observed behavior on benchmark-v1.0. Exactly 8 cases (1.6%) exhibited minor regressions (primarily brevity in formal mathematical step derivations and concise logic proofs).

3. Quickstart (Inference Example)

python

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

base_model_id = "Qwen/Qwen3-4B-Base"
adapter_dir = "releases/qwen3-4b-qlora-v1.0/adapter"

# 1. Configure 4-bit NF4 Quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)

# 2. Load Base Model and Tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)

# 3. Mount LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_dir)
model.eval()

# 4. Generate Response using ChatML Template
messages = [
    {"role": "system", "content": "You are a helpful, precision-driven technical assistant."},
    {"role": "user", "content": "Write a Python context manager for atomic file writing."},
]
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=512,
        temperature=0.0,
        do_sample=False,
    )
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

4. Cryptographic Release Verification

To verify the tamper-evident integrity and reproducibility of all release artifacts:

python scripts/validate_final_release.py --release releases/qwen3-4b-qlora-v1.0

durgeshkidsyt/qwen3-4b-qlora-v1.0