phi4-3b-ec-magento

phi4-3b-ec-magento

A QLoRA adapter for microsoft/Phi-4-mini-instruct (3.8B) specialized in e-commerce catalog tasks with Magento 2 conventions: attribute extraction to JSON, product Q&A, search-relevance classification, and product ranking. Third model in a same-data, same-recipe, same-harness comparison with qwen3.5-4b-ec-magento and ministral3-3b-ec-magento.

The adapter was trained on unsloth/Phi-4-mini-instruct-bnb-4bit (a standard NF4 export of the microsoft weights, with the eos fix — see gotchas); the GGUF is the adapter merged into the original bf16 weights.

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(
    "gabrielgts/phi4-3b-ec-magento", max_seq_length=2048, load_in_4bit=True)
FastLanguageModel.for_inference(model)

# No system message — the model was trained without one (Phi's template
# injects nothing by default). Greedy decoding recommended.
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"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text=text, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
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, "gabrielgts/phi4-3b-ec-magento").

Usage — GGUF / Ollama

hf download gabrielgts/phi4-3b-ec-magento --include 'gguf/*' --local-dir .
cd gguf && ollama create phi4-3b-ec-magento -f Modelfile
ollama run phi4-3b-ec-magento "target attribute: color ..."

Gotchas (learned the hard way)

• Fine-tune from the unsloth repo, not microsoft's: microsoft's tokenizer config sets eos=<|endoftext|>, which the model rarely emits — generation does not stop at <|end|> and rambles past answers. The unsloth export fixes eos=<|end|>. Unsloth also silently remaps the microsoft id to its dynamic-quant repo; pin the plain -bnb-4bit id for standard NF4.
• Fused projections: phi3 fuses qkv_proj/gate_up_proj. LoRA target names q_proj/k_proj/v_proj/gate_proj/up_proj silently match nothing (peft skips without erroring) — target qkv_proj, o_proj, gate_up_proj, down_proj.
• Quantization sensitivity: at plain Q4_K_M the tied 200k-vocab embedding degrades outputs (wrong attributes extracted with perfect JSON formatting). The shipped GGUF uses --token-embedding-type q8_0 --output-tensor-type q8_0.

Training recipe

Evaluation

Greedy decoding, identical prompts across all models; base models evaluated zero-shot with the same harness. Siblings: qwen3.5-4b-ec-magento (4.55B params), ministral3-3b-ec-magento (3.85B text params) — Phi-4-mini (3.84B) and Ministral are size-matched; Qwen is ~18% larger.

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

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

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; they understate its underlying capability.
• The fused-projection LoRA shares one rank-8 subspace across q/k/v — a small ECInstruct deficit vs the siblings may reflect this structural difference rather than model quality.
• English only; fine-tuned on structured data — expect degraded general chat vs the base (drop the adapter to recover it).
• Use greedy decoding (do_sample=False / temperature 0) — that's how it was evaluated.

Provenance

References

• Phi-4-mini: arXiv:2503.01743
• ECInstruct / eCeLLM: arXiv:2402.08831
• QLoRA: arXiv:2305.14314