qwen3.5-4b-ec-magento

A QLoRA adapter for Qwen/Qwen3.5-4B specialized in e-commerce catalog tasks with Magento 2 conventions: attribute extraction to JSON, product Q&A, search-relevance classification, and product ranking.

The repo contains both distribution formats:

Path	What	Size	For
/ (root)	LoRA adapter (PEFT)	42MB	transformers/peft/unsloth on top of the base model
gguf/	Merged + quantized Q4_K_M + Ollama Modelfile	2.6GB	llama.cpp / Ollama, CPU+GPU serving

What it does

Trained on 56,253 instruction samples: 70% ECInstruct (generic e-commerce) + 30% synthetic Magento-schema data generated from the Magento Luma sample catalog (products fully disjoint between train and eval). Four task shapes:

Attribute extraction — product text (or a raw Magento custom_attributes payload) → JSON:

target attribute: size
product title: Puma Suede green sneakers size 43
→ [{"attribute": "size", "value": "43"}]

Absent attributes are reported as "None" rather than hallucinated.

Product QA — a question answered strictly from given product data. Relevance classification — query + product → graded relevance option (ESCI-style A–D). Relevance ranking — query + lettered product list → ranked letters (B,A,C).

Usage — adapter (unsloth / peft)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "gtstadeu/qwen3.5-4b-ec-magento", max_seq_length=2048, load_in_4bit=True)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content":
    "Extract the value of the target attribute from the given product information "
    "and output it as JSON. If the attribute is not present, output None as the value.\n\n"
    "target attribute: size\nproduct title: Puma Suede green sneakers size 43"}]

# Qwen3.5 gotchas: enable_thinking=False is REQUIRED (the adapter was trained with a
# closed <think> block; an open one makes it emit reasoning instead of the answer),
# and the multimodal Processor requires tokenizer(text=...), never positional.
text = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, enable_thinking=False, tokenize=False)
inputs = tokenizer(text=text, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=64, do_sample=False)  # greedy recommended
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# [{"attribute": "size", "value": "43"}]

Plain peft also works: PeftModel.from_pretrained(base_model, "gtstadeu/qwen3.5-4b-ec-magento") over Qwen/Qwen3.5-4B.

Usage — GGUF / Ollama

hf download gtstadeu/qwen3.5-4b-ec-magento --include 'gguf/*' --local-dir .
cd gguf && ollama create qwen3.5-4b-ec-magento -f Modelfile
ollama run qwen3.5-4b-ec-magento --think=false "target attribute: color ..."

Ollama ≥ 0.31 uses its built-in Qwen3.5 renderer and defaults to thinking mode — pass --think=false (CLI) or "think": false (API) for the trained fast-path output. Requires an Ollama version with qwen35 architecture support. ~95 tok/s on an RTX 3070.

Training recipe

Parameter	Value
Method	QLoRA (4-bit NF4 base, bf16 compute) via Unsloth
LoRA	r=8, alpha=16, targets q/k/v/o/gate/up/down_proj
Trainable params	10.6M of 4,550M (0.23%)
Batch	1 × grad_accum 16 (effective 16), max_seq_length 2048
Optimizer / LR	paged_adamw_8bit, 2e-4 cosine, 1 epoch, seed 42
Data	56,253 samples: 39,377 ECInstruct + 16,876 Magento-synthetic

Evaluation

Greedy decoding, identical prompts and chat template across all models; base model evaluated zero-shot with the same harness. ec50k is an intermediate adapter trained on ECInstruct only (not published) — shown to isolate what the Magento data adds.

Magento held-out set (2,969 samples, 475 products never seen in training):

Task · metric	Base	ec50k	this model
Attribute extraction · F1	0.015	0.320	0.938
Attribute extraction · parse failures	71.5%	8.1%	0%
Product QA · token-F1	0.041	0.299	0.943
Relevance classification · accuracy	0.194	0.770	0.962
Relevance rank · top-1	0.345	0.770	0.799

ECInstruct held-out set (2,000 samples):

Task · metric	Base	ec50k	this model
Attribute extraction · F1	0.000	0.685	0.646
Query→product rank · top-1	0.013	0.670	0.650
Relevance classification · accuracy	0.258	0.678	0.685
Answerability · accuracy	0.503	0.778	0.780

Limitations — read before relying on the numbers

• The Magento eval is synthetic-on-synthetic. Eval tasks were generated with the same templates as training data (products fully disjoint). It validly measures schema adherence — JSON format, Magento attribute vocabularies, the None-when-absent rule — but overstates production quality on real catalogs and real user queries.
• The base model's near-zero extraction scores are dominated by format non-adherence (it answers in prose); they understate its underlying capability, though prose output is itself a blocker for programmatic use.
• English only. Inputs should fit a 2,048-token budget (strip HTML from product descriptions).
• Free-form generation is weak (trained r=8, extraction-focused); use it for structured tasks, not copywriting.
• Use greedy decoding (do_sample=False / temperature 0) — that's how it was evaluated.

Provenance

Training run	qwen3.5-4b-r8-mix56k-e1 (2026-07-03)
Adapter sha256	8e7a7f89da314bd7c206b1f2fad2ad9eaaa4b0e988f72ea8f4592aa099bc6b42
Train set sha256	afb7e664cda4490597c1914db6aff94933991a56d2175435666f8f6b7a726532 (mixture_train.jsonl, 56,253 rows)
Eval set sha256	8feafdb2… (magento_eval.jsonl) · 2583a61e… (ecinstruct_eval.jsonl)
Stack	torch 2.6.0+cu124 · transformers 5.5.0 · unsloth 2026.6.9

References

• ECInstruct / eCeLLM: arXiv:2402.08831
• QLoRA: arXiv:2305.14314
• LoRA rank for merchant extraction (r=8 finding): arXiv:2606.08051