RetailOps-Instruct Qwen3.5-9B

RetailOps-Instruct is a full merged Transformers checkpoint for e-commerce catalog operations.

It is built by merging:

• Base model: Qwen/Qwen3.5-9B
• LoRA adapter: kyLELEng/retailops-instruct-qwen3.5-9b-lora

No additional training is performed during this merge step.

Task

The model is intended to turn product metadata, category information, seller notes, and customer reviews into structured listing packages:

• optimized_title
• bullet_points
• description
• attributes
• seo_keywords
• faq
• review_driven_improvements
• compliance_notes

Training Data

• SFT dataset: kyLELEng/retailops-instruct-sft
• Sources in v1:
  • AI4H/EC-Guide
  • McAuley-Lab/Amazon-Reviews-2023
  • McAuley-Lab/Amazon-C4

Training Summary

• Method: LoRA SFT, then merged into the base model
• LoRA rank: 64
• LoRA alpha: 128
• LoRA dropout: 0.05
• Max sequence length: 2048

Evaluation

Final adapter eval before merge:

{
  "json_validity_rate": 0.5,
  "avg_required_field_completion": 0.171875,
  "num_generation_eval_samples": 8
}

No-thinking generation check:

{
  "inference_mode": "tokenizer.apply_chat_template(..., enable_thinking=False) when supported",
  "num_generation_eval_samples": 24,
  "json_validity_rate": 1.0,
  "listing_num_samples": 13,
  "listing_json_validity_rate": 1.0,
  "listing_avg_required_field_completion": 0.7980769230769231
}

Recommended Inference

Qwen3.5 models may emit reasoning text by default. For this RetailOps model, use no-thinking mode when available:

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

Example

import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "kyLELEng/retailops-instruct-qwen3.5-9b"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are RetailOps-Instruct, an e-commerce catalog optimization assistant. Return exactly one valid JSON object. Do not invent unsupported product features. Use product metadata and reviews as the source of truth."},
    {"role": "user", "content": "Given the product metadata, category, and customer reviews, generate an optimized product listing package.\n\nINPUT:\n" + json.dumps({
        "category": "Home & Kitchen > Kitchen & Dining > Small Appliances",
        "brand": "ExampleBrand",
        "raw_title": "Portable Electric Kettle",
        "raw_description": "Small kettle for boiling water.",
        "product_specs": {"capacity": "1.0L", "material": "stainless steel", "safety": "auto shut-off"},
        "positive_reviews": ["Heats water quickly.", "Good size for small apartments."],
        "negative_reviews": ["The instructions were unclear.", "The outside gets warm after use."],
        "brand_voice": "clear, practical, trustworthy"
    }, ensure_ascii=False, indent=2)},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=900, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Merge Smoke Test

{
  "input": {
    "category": "Home & Kitchen > Kitchen & Dining > Small Appliances",
    "brand": "ExampleBrand",
    "raw_title": "Portable Electric Kettle",
    "raw_description": "Small kettle for boiling water.",
    "product_specs": {
      "capacity": "1.0L",
      "material": "stainless steel",
      "safety": "auto shut-off"
    },
    "positive_reviews": [
      "Heats water quickly.",
      "Good size for small apartments."
    ],
    "negative_reviews": [
      "The instructions were unclear.",
      "The outside gets warm after use."
    ],
    "brand_voice": "clear, practical, trustworthy"
  },
  "generated": "{\"optimized_title\": \"Portable Electric Kettle\", \"bullet_points\": [\"Small kettle for boiling water.\"], \"description\": \"Small kettle for boiling water.\", \"seo_keywords\": [\"kettle\", \"small\", \"portable\", \"electric\", \"boiling\", \"water\", \"examplebrand\", \"home\", \"kitchen\", \"dining\"], \"attributes\": {\"brand\": \"ExampleBrand\", \"category\": \"Home & Kitchen > Kitchen & Dining > Small Appliances\", \"features\": [\"stainless steel\", \"auto shut-off\"], \"specifications\": {\"capacity\": \"1.0L\", \"material\": \"stainless steel\", \"safety\": \"auto shut-off\"}}, \"review_driven_improvements\": [\"Highlight review-supported strengths: water, heats, quickly, good, size\", \"Clarify or avoid overclaiming around reported issues: instructions, were, unclear, outside, gets\", \"Do not contradict the supplied customer review evidence.\"], \"faq\": [{\"question\": \"Who is the brand?\", \"answer\": \"ExampleBrand\"}, {\"question\": \"What category does this product fit?\", \"answer\": \"Home & Kitchen > Kitchen & Dining > Small Appliances\"}], \"compliance_notes\": [\"No unsupported product features added.\", \"No medical, safety, or durability claims added unless present in the input.\"]}"
}

Limitations

• The model can still produce unsupported claims; downstream validation is recommended.
• This is not a compliance certification system.
• Outputs should be checked before production publishing.