ministral3-3b-ec-magento

ministral3-3b-ec-magento

A QLoRA adapter for Ministral 3 3B Instruct (2512) specialized in e-commerce catalog tasks with Magento 2 conventions: attribute extraction to JSON, product Q&A, search-relevance classification, and product ranking. Sibling model to gtstadeu/qwen3.5-4b-ec-magento — same frozen data, same recipe, same eval harness — for cross-family comparison.

The adapter was trained on unsloth/Ministral-3-3B-Instruct-2512-bnb-4bit (a standard NF4 export of mistralai/Ministral-3-3B-Instruct-2512); 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/ministral3-3b-ec-magento", max_seq_length=2048, load_in_4bit=True)
FastLanguageModel.for_inference(model)

# An EXPLICIT EMPTY system message is REQUIRED: without one, the Ministral 3
# chat template injects Mistral's ~520-token default system prompt, which the
# adapter was not trained with. The multimodal Processor also requires
# tokenizer(text=...), never a positional string.
messages = [
    {"role": "system", "content": ""},
    {"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)  # 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, "gabrielgts/ministral3-3b-ec-magento").

Usage — GGUF / Ollama

hf download gabrielgts/ministral3-3b-ec-magento --include 'gguf/*' --local-dir .
cd gguf && ollama create ministral3-3b-ec-magento -f Modelfile
ollama run ministral3-3b-ec-magento "target attribute: color ..."

The bundled Modelfile renders the trained prompt format ([SYSTEM_PROMPT][/SYSTEM_PROMPT][INST]…[/INST]) with temperature 0. If your Ollama version ships a built-in Ministral 3 renderer, pass an explicit empty system prompt per request so Mistral's default one is not injected.

Training recipe

Evaluation

Greedy decoding, identical prompts and chat template across all models; base model evaluated zero-shot with the same harness. The Qwen column is the same-data, same-recipe sibling qwen3.5-4b-ec-magento (4.55B params vs 3.85B text params here — mind the ~15% size gap when comparing).

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 (it answers in prose); they understate its underlying capability, though prose output is itself a blocker for programmatic use.
• Fine-tuned on English-only structured data: expect degraded multilingual and general-chat ability versus the base model (drop the adapter to recover it). The vision tower is untouched, but this repo's GGUF is text-only.
• 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

References

• Ministral 3: arXiv:2601.08584
• ECInstruct / eCeLLM: arXiv:2402.08831
• QLoRA: arXiv:2305.14314