Model Card for zeroproof-ecommerce-0.5b

E-commerce intent detection in a 0.5B model, the smallest member of the ZeroProof family. Given a live customer conversation, zeroproof-ecommerce-0.5b returns a single JSON object with the customer's payment intent and its structured details, at the lowest cost and latency in the family.

On a hard, held-out benchmark it reaches 65.5 macro-averaged intent-type accuracy, roughly 94% of our 1B, from a base model that scores 10.7.

Model Details

Model Description

• Developed by: ZeroProof
• Model type: E-commerce payment-intent classifier; structured JSON output over seven intent types
• Language: English
• License: Apache-2.0 (inherited from the base model)
• Finetuned from: Qwen/Qwen2.5-0.5B-Instruct, via 4-bit QLoRA

Model Sources

• Repository: https://huggingface.co/zero-proof-ai/zeroproof-ecommerce-0.5b
• Training data: https://huggingface.co/datasets/zero-proof-ai/ecommerce-intent

Uses

Direct Use

Drop-in intent detection for e-commerce and payments support. Feed the conversation so far; the model returns one JSON object per message: whether an actionable payment intent is present, which of seven types it is, and the extracted details. The smallest, cheapest member of the family, built to run inline on every turn.

Out-of-Scope Use

English-language e-commerce and payments only. It is a narrow intent classifier, not a general assistant, and should not be the sole authority for executing a payment without a downstream verification step.

Bias, Risks, and Limitations

Trained and evaluated on synthetic role-play conversations from a multi-agent marketplace simulation, labeled in a separate model pass under a locked labeling policy; labels are model-assigned and carry residual noise, and live-traffic distribution may differ. Accuracy is strongest on subscriptions, refunds, and bill-pay; asset-exchange and checkout are the focus of the current data round. It also holds the no-action boundary more loosely than the 1B (13.5% vs 4.2% of true no-action conversations marked actionable); prefer the 1B where a false positive is costly.

Recommendations

Pair the model with a verification layer before any payment executes, and re-benchmark on your own traffic before relying on the reported numbers in production.

How to Get Started with the Model

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained("zero-proof-ai/zeroproof-ecommerce-0.5b")
tok   = AutoTokenizer.from_pretrained("zero-proof-ai/zeroproof-ecommerce-0.5b")

Use greedy decoding (temperature=0). The system prompt and per-intent detail schemas ship in the repo.

Training Details

Training Data

The differentiator is the data. ZeroProof builds e-commerce intent training data as randomized conversational simulations: role-played customers with independently sampled personas, situations, and behaviors including adversarial actors, simulated turn by turn between two independently drawn models. Generation is label-blind, labels are assigned in a separate pass, and every candidate passes a structural data gate with zero train/eval leakage. This 0.5B was trained on the same class-rebalanced split as the 1B so it routes across all intent types rather than defaulting to the majority.

Training Procedure

Training Hyperparameters

• Method: 4-bit QLoRA (via unsloth)
• Epochs: 1
• Base: Qwen2.5-0.5B-Instruct

Evaluation

Testing Data, Factors & Metrics

Held-out eval of 1,977 conversations, zero train/eval leakage, macro-averaged, scored deterministically against gold: intent type by exact match, order details as the fraction of gold fields reproduced (partial credit per field). The frontier panel is scored on a 412-row balanced subset of the same eval (reused, not re-run).

Results

• Fine-tuning takes intent-type accuracy from 10.7 to 65.5 (macro), about 6x the base, at ~94% of our 1B and a fraction of frontier size.
• On structured detail extraction the 0.5B beats every frontier panel member: 57.6 against Sonnet 5's 55.7, Opus 4.8's 53.2, and GPT-5's 41.4. The 1B leads the family at 63.2.
• The smallest member of the family, built to run inline on live traffic at the lowest cost and latency.

Technical Specifications

Model Architecture and Objective

A LoRA adapter over Qwen2.5-0.5B-Instruct, trained response-only to emit one structured intent object per turn.

Model Card Contact

ZeroProof, https://huggingface.co/zero-proof-ai