toolathon123/logistics-southamerica-quality

Logistics Quality – South America (Brazil & Argentina)

Live shipment records collected from the South American (Brazil / Argentina) logistics network for service-quality monitoring and improvement. This dataset supports complaint type classification of customer feedback, on-time/damage KPI reporting, and route-level hotspot analysis (with special attention to cross-border corridors such as São Paulo → Buenos Aires).

Collection & real-time nature

Records are ingested from the transport management system in near real-time. Each snapshot is timestamped in UTC. New records can be appended incrementally — see Incremental append — without re-processing historical rows.

Schema (columns)

Column	Type	Description
transport_id	string	Unique shipment id (e.g. BR-20260813-000001)
timestamp	string	ISO-8601 UTC timestamp of the shipment event
origin_country	string	BR or AR
origin_city	string	Origin city
destination_country	string	BR or AR
destination_city	string	Destination city
route_key	string	"{origin} -> {destination}" corridor key
is_cross_border	bool	Whether the shipment crosses the BR–AR border
transport_mode	string	road | rail | air
carrier_name	string	Carrier operating the shipment
on_time_delivery	bool	Delivered within agreed deadline
delay_hours	float	Delay in hours (0 when on time)
damage_status	string	none | minor | major
feedback_language	string	Language of the customer feedback (pt / es)
customer_feedback	string	Free-text customer feedback (Portuguese or Spanish)
gt_complaint	string	Reference complaint label used for validation: none|delay|damage|service|cost|loss
predicted_complaint	string	Model-predicted complaint category (see models below)
confidence	float	Confidence of the top predicted category
prob_delay … prob_positive	float	Full model probability vector over the 6 categories
segment	string	Ingestion batch (live_batch, incremental_*)

Complaint categories

Category	Meaning
delay	延误 — late / delayed delivery
damage	货损 — damaged or spoiled goods
service	服务态度 — poor service / attitude
cost	费用争议 — billing / cost dispute
loss	包裹丢失 — lost / missing package
positive	no complaint / positive feedback

Models used for inference

• Primary: MoritzLaurer/mDeBERTa-v3-base-mnli-xnli — multilingual zero-shot NLI (XLM-R/DeBERTa family). One model handles both Portuguese (BR) and Spanish (AR) feedback via the Hugging Face Inference API.
• Cross-check: joeddav/xlm-roberta-large-xnli — XLM-RoBERTa-large NLI, used to validate predictions on a sample.
• Language-specific pretrained models considered: neuralmind/bert-base-portuguese-cased (BERTimbau, pt) and dccuchile/bert-base-spanish-wwm-cased (BETO, es).

Incremental append

from datasets import load_dataset
from huggingface_hub import HfApi

api = HfApi()
# append a small batch of new records (only new transport_ids)
new_records = [...]  # list of dicts with the schema above (new transport_ids)
api.upload_file(
    path_or_fileobj=... ,  # or use push_to_hub on the dataset
    path_in_repo="data/incremental_20260813.parquet",
    repo_id="toolathon123/logistics-southamerica-quality",
    repo_type="dataset",
)
# Re-run inference only on records lacking a `predicted_complaint`.

A fully worked append + re-classification example is provided in the analysis pipeline (see logistics_project/scripts/).