{
  "productName": "Apple iPhone 14 Pro 128GB Space Black",
  "brand": "Apple",
  "model": "iPhone 14 Pro",
  "condition": "neuwertig",
  "category": "elektronik",
  "attributes": {"storage": "128GB", "color": "Space Black"}
}

Training Details

Base model: unsloth/Qwen3.5-2B
Method: bf16 LoRA with unsloth (QLoRA not recommended for Qwen3.5)
LoRA rank: 16
LoRA alpha: 32
Training samples: 4002
Eval samples: 445
Final loss: 0.7330062726579339
Epochs: 3

Usage with vLLM

python

from vllm import LLM, SamplingParams

llm = LLM(model="ekwav/ane-extraction-qwen3-1.7b", dtype="float16")
params = SamplingParams(temperature=0.0, max_tokens=256, stop=["<|im_end|>"])

prompt = '''<|im_start|>system
You are a JSON extractor for marketplace listings. Output ONLY a valid JSON object.<|im_end|>
<|im_start|>user
T:iPhone 14 Pro 128GB
D:Neuwertig, immer mit Hülle benutzt
P:749EUR /no_think<|im_end|>
<|im_start|>assistant
'''

output = llm.generate([prompt], params)
print(output[0].outputs[0].text)

Intended Use

Part of the ANE (ane.deals) product extraction pipeline. Designed to replace larger models for cost-efficient inference at scale.

Ekwav/ane-extraction-qwen3.5-2b

ANE Extraction Model — Qwen3.5 2B (Merged)

Model Description

Training Details

Usage with vLLM

Intended Use

Ekwav/
ane-extraction-qwen3.5-2b