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

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

This model is presented in the paper ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification.

Atlas is a domain-specialized LLaMA-3.3-70B model fine-tuned on U.S. Customs CROSS rulings for Harmonized Tariff Schedule (HTS) code assignment.
It targets both 10-digit U.S. HTS (compliance) and 6-digit HS (globally harmonized) accuracy.

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

Atlas outperforms general-purpose LLMs while remaining deployable/self-hostable.

• Model repo: flexifyai/atlas-llama3.3-70b-hts-classification
• Dataset: flexifyai/cross_rulings_hts_dataset_for_tariffs
• Demo: flexifyai/atlas-llama3_3-70b-hts-demo
• Project page: https://tariffpro.flexify.ai/

Example (from the demo):

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 -> Falls under heterocyclic compounds with nitrogen hetero-atom(s); specifically classified within pteridine derivatives used in pharmaceutical or biochemical applications per CROSS rulings.

---

TL;DR

• Task: Assign an HTS code given a product description (and optionally rationale).
• Why it matters: Misclassification halts shipments; 6-digit HS is global, 10-digit is U.S.-specific.
• What’s new: First open benchmark + strong open model baseline focused on semiconductors/manufacturing.

---

Intended use & limitations

Use cases

• Automated HTS/HS pre-classification with human-in-the-loop review.
• Decision support for brokers, compliance, and trade workflows.
• Research on domain reasoning, retrieval, and alignment.

Limitations

• Not legal advice; rulings change and are context-dependent.
• Training data is concentrated in semiconductors/manufacturing; performance may vary elsewhere.
• Model can produce confident but incorrect codes; keep a human validator for high-stakes usage.
• Always verify against the current HTS/USITC and local customs guidance.

---

Data

• Source: CROSS (U.S. Customs Rulings Online Search System).
• Splits: 18,254 train / 200 valid / 200 test.
• Each example includes:
  • product description
  • chain-of-reasoning style justification
  • ground-truth HTS code

Dataset card: flexifyai/cross_rulings_hts_dataset_for_tariffs

---

Training setup (summary)

• Base: LLaMA-3.3-70B (dense)
• Objective: Supervised fine-tuning (token-level NLL)
• Optimizer: AdamW (β1=0.9, β2=0.95, wd=0.1), cosine LR schedule, peak LR 1e-7
• Precision: bf16, gradient accumulation (effective batch ≈ 64 seqs)
• Hardware: 16× A100-80GB, 5 epochs (~1.4k steps)

We chose a dense model for simpler finetuning/inference and reproducibility under budget constraints.
Future work: retrieval, DPO/GRPO, and smaller distilled variants.

---

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 on A100s can be significantly cheaper per 1k inferences than proprietary APIs.

---

Prompting

Atlas expects an instruction like:

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, un-doped, for semiconductor fabrication?

Model:
HTS US Code -> 3818.00.0000
Reasoning -> Classified under chemical elements/compounds doped for electronics; wafer form per CROSS precedents.

---

Authors

• Pritish Yuvraj (Flexify.AI) — pritishyuvraj.com
• Siva Devarakonda (Flexify.AI)

📖 Citation

If you find this work useful, please cite our paper:

@misc{yuvraj2025atlasbenchmarkingadaptingllms,
  title={ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification}, 
  author={Pritish Yuvraj and Siva Devarakonda},
  year={2025},
  eprint={2509.18400},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2509.18400}, 
}