Request access to ATLAS (Flexify.AI Inc.)

Your request will be manually reviewed. Please confirm attribution requirements below.

By requesting access, you agree to provide visible attribution in any public use, hosting, derivative model card, benchmark, publication, or demo: “ATLAS HTS Classification Model by Flexify.AI Inc. (https://www.flexify.ai)”. For commercial or production usage without public attribution, please contact [email protected].

Atlas — LLaMA-3.3-70B fine-tuned for Harmonized Tariff Schedule (HTS) classification

Atlas is a fine-tuned LLaMA-3.3-70B model for automated Harmonized Tariff Schedule (HTS) and Harmonized System (HS) code classification, designed for trade compliance and manufacturing workflows.

Maintained by Flexify.AI Inc. as part of the ATLAS trade intelligence research program.

Atlas is trained on U.S. Customs CROSS rulings and targets both:

10-digit U.S. HTS (compliance-critical)
6-digit HS (globally harmonized)

Evaluation highlights

10-digit exact match: 40.0%
6-digit exact match: 57.5%

Atlas significantly outperforms general-purpose LLMs on HTS classification while remaining self-hostable and suitable for enterprise deployment.

Resources

Model repo: https://huggingface.co/flexifyai/atlas-llama3.3-70b-hts-classification
Dataset: https://huggingface.co/datasets/flexifyai/cross_rulings_hts_dataset_for_tariffs
Demo: https://flexifyai-atlas-llama3-3-70b-hts-demo.hf.space
Project page: https://tariffpro.flexify.ai/

Research & Community Visibility

This model is described in the paper:
ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification
https://huggingface.co/papers/2509.18400

The ATLAS methodology and results were also presented at:

NeurIPS 2025 - NORA Knowledge Graphs Workshop
Session: ATLAS and HTS LLMs for Global Trade
https://neurips.cc/virtual/2025/loc/mexico-city/129845

Example

User:
What is the HTS US Code for 4[N-(2,4-Diamino-6-Pteridinylmethyl)-N-Methylamino] Benzoic Acid Sodium Salt?

Model:
HTS US Code → 2933.59.4700
Reasoning → Classified under heterocyclic compounds with nitrogen hetero-atoms; specifically pteridine derivatives used in pharmaceutical or biochemical applications per CROSS rulings.

TL;DR

Task: Assign an HTS/HS code given a product description
Why it matters: Misclassification can halt shipments or trigger penalties
Scope: 6-digit HS (global) and 10-digit HTS (U.S.-specific)
What’s new: A focused benchmark and strong open baseline for trade and manufacturing domains

Intended Use & Limitations

Intended Use

Automated HTS/HS pre-classification with human-in-the-loop review
Decision support for customs brokers, compliance, and trade teams
Research on domain-specific reasoning and structured classification

Limitations

Not legal advice; tariff rulings change and are context-dependent
Training data is concentrated in manufacturing and semiconductors
The model may produce confident but incorrect codes
Always validate against current USITC HTS and local customs guidance

Risks & Considerations

The model reflects historical ruling patterns and may encode biases present in prior decisions
Performance may degrade for products or industries under-represented in CROSS
Not suitable for autonomous compliance decisions without expert oversight

Data

Source: U.S. Customs Rulings Online Search System (CROSS)
Splits: 18,254 train / 200 validation / 200 test
Each example includes:
- Product description
- Reasoning-style justification
- Ground-truth HTS code

Dataset card:
https://huggingface.co/datasets/flexifyai/cross_rulings_hts_dataset_for_tariffs

Training Summary

Base model: LLaMA-3.3-70B (dense)
Objective: Supervised fine-tuning (token-level NLL)
Optimizer: AdamW (β1=0.9, β2=0.95, weight decay=0.1)
Schedule: Cosine LR, peak LR 1e-7
Precision: bf16
Hardware: 16× A100-80GB
Epochs: 5 (~1.4k steps)

A dense model was chosen for reproducibility and simpler deployment.
Future work: retrieval augmentation, preference optimization, and distilled variants.

Evaluation Results (200-example held-out test)

Model	10-digit exact	6-digit exact	Avg. digits correct
GPT-5-Thinking	25.0%	55.5%	5.61
Gemini-2.5-Pro-Thinking	13.5%	31.0%	2.92
DeepSeek-R1 (05/28)	2.5%	26.5%	3.24
GPT-OSS-120B	1.5%	8.0%	2.58
LLaMA-3.3-70B (base)	2.1%	20.7%	3.31
Atlas (this model)	40.0%	57.5%	6.30

💰 Cost note: Self-hosting Atlas can be significantly cheaper per 1k inferences than proprietary APIs.

Prompting

Expected format: User: What is the HTS US Code for [product_description]? Model: HTS US Code -> [10-digit code] Reasoning -> [short justification]

Minimal Example

User:
What is the HTS US Code for 300mm silicon wafers, polished, undoped, for semiconductor fabrication?

Model:
HTS US Code → 3818.00.0000
Reasoning → Classified under chemical elements and compounds for electronics; wafer form per CROSS precedents.

Maintained By

Siva Devarakonda — Founder & CEO, Flexify.AI Inc.

📖 Citation

Model

@model{flexify_atlas_hts_llama,
  title={ATLAS HTS Classification Model (LLaMA-3.3-70B)},
  author={{Flexify.AI Inc.}},
  year={2025},
  url={https://huggingface.co/flexifyai/atlas-llama3.3-70b-hts-classification},
  howpublished={\url{https://www.flexify.ai}}
}

@misc{atlas2025benchmarking,
  title={ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification},
  author={Siva Devarakonda},
  year={2025},
  eprint={2509.18400},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2509.18400}
}

Downloads last month: 21

Safetensors

Model size

71B params

Tensor type

BF16

Model tree for flexifyai/atlas-llama3.3-70b-hts-classification

Base model

meta-llama/Llama-3.1-70B

Finetuned

meta-llama/Llama-3.3-70B-Instruct

Finetuned

(258)

this model

Dataset used to train flexifyai/atlas-llama3.3-70b-hts-classification

Paper for flexifyai/atlas-llama3.3-70b-hts-classification

ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification

Paper • 2509.18400 • Published Sep 22, 2025 • 4