xlm-roberta-base-text-geolocation

FacebookAI/xlm-roberta-base fine-tuned end-to-end for multiclass classification of short social-media text into one of 123 geographic regions. Output is a region label with a probability distribution; the model does not predict coordinates.

Labels

yachay/text_coordinates_regions contains one JSON file per region (c_0.json … c_122.json) and no region names. Labels are these file ids (c_0 … c_122). For orientation, the table at the end lists the empirical medoid of each region's training coordinates (descriptive statistic, not an official region definition).

Data

• Source: yachay/text_coordinates_regions @ b9fa48181e3c93791d0613382c77aeb91214f324, 615,000 posts, 5,000 per region (balanced). Coordinates are place-level (≈12k distinct points).
• Model input: text only. Coordinates are used for evaluation only.
• Preprocessing: strip, lower-case (corpus is already lower-cased), URLs → HTTPURL. Mentions, hashtags, emoji and place names are kept.

Filter	Removed	Reason
null_or_empty	0	No text or no label: unusable.
malformed_coordinates	0	Coordinates not two numbers in valid range.
url_only	12,911	Text consists only of t.co links; the hash has no textual content.
duplicate_text_region	1,286	Identical text with identical label: keep one copy.

Remaining: 600,803 posts (2.31% removed).

• Split: 480,496 / 60,103 / 60,204 (train / validation / test), stratified by region, seed 42.
• Leakage control: posts that are identical after removing URLs and collapsing whitespace form a group (4,776 multi-post groups) and are assigned to a single split. Exact-text and group overlap between splits: 0.

Training

Parameter	Value
Base model	FacebookAI/xlm-roberta-base
Head	linear classification head (XLMRobertaForSequenceClassification)
Loss	cross-entropy (no class weighting; classes are balanced)
Max sequence length	256 (99.95% of posts are ≤ 256 tokens)
Optimizer	AdamW (fused), weight decay 0.01
Learning rate	2e-05, linear schedule, warmup ratio 0.1
Epochs	3.0
Batch size	32 × 1 accumulation = 32
Gradient clipping	1.0
Precision	bf16
Padding	dynamic
Model selection	best validation macro_f1 (evaluated each epoch)
Seed	42
Hardware	NVIDIA GeForce RTX 5060 Ti
Training time	1.27 h
Software	torch 2.11.0+cu128, transformers 5.17.0, datasets 4.3.0

Evaluation

Classification (60,204 test posts):

Metric	Validation	Test	TF-IDF + LR (test)
Accuracy (top-1)	24.56	24.34	27.35
Top-3 accuracy	44.88	45.04	45.89
Top-5 accuracy	56.11	56.12	55.76
Macro F1	23.21	22.95	26.81
Weighted F1	23.13	22.86	26.72
Macro precision	25.80	25.23	27.51
Macro recall	24.63	24.43	27.44

Geographic distance (test). Distance between the post's coordinate and the medoid of the predicted region. "Oracle" uses the medoid of the true region, i.e. the floor for a perfect classifier under this representation.

Metric	Model	Oracle
Median distance to predicted region medoid (km)	1264	181
Mean distance (km)	3126	272
Acc@161 km	12.73	46.54
Acc@500 km	27.27	87.62
Acc@1000 km	44.01	96.42
Acc@2500 km	66.68	99.46

Usage

from transformers import pipeline

clf = pipeline("text-classification", model="thealper2/xlm-roberta-base-text-geolocation", top_k=5)
print(clf("just landed at jfk, the traffic in queens is already insane"))

Apply the training preprocessing (lower-case, URLs → HTTPURL) for best results; settings are stored in preprocessing.json.

Limitations

• Region labels are unnamed dataset clusters; regions have unequal geographic extent.
• Coordinates in the source data are place-level, so distance metrics are coarse.
• Twitter data from 2021: topical, demographic and platform biases; performance on other domains or periods is not measured.
• Many posts carry no geographic signal (emoji-only, generic replies); confidence is low for these.
• Some posts contain explicit place names from app templates (check-ins, "just posted a photo @ …"), which are easy cases.

Per-region results (test)

Region	Medoid lat	Medoid lon	Median km to medoid	F1
c_0	40.56	-75.06	251	0.055
c_1	52.72	-1.35	198	0.121
c_2	35.61	137.77	189	0.356
c_3	-22.29	-43.55	257	0.035
c_4	34.59	-83.21	309	0.027
c_5	40.24	-3.53	357	0.180
c_6	33.96	-117.85	124	0.045
c_7	41.61	-85.25	242	0.017
c_8	-28.24	-50.83	272	0.086
c_9	28.73	76.83	234	0.095
c_10	45.71	9.43	280	0.036
c_11	34.02	-95.64	270	0.034
c_12	39.71	29.71	258	0.355
c_13	-6.53	107.85	201	0.291
c_14	-34.72	-58.92	297	0.150
c_15	27.57	-81.03	151	0.101
c_16	19.43	-98.79	121	0.163
c_17	24.99	47.32	285	0.266
c_18	3.22	101.96	131	0.421
c_19	14.57	120.94	110	0.377
c_20	4.84	-74.15	142	0.188
c_21	30.25	-91.74	273	0.122
c_22	33.47	130.74	120	0.414
c_23	26.55	82.89	218	0.338
c_24	19.52	73.14	186	0.234
c_25	29.69	-98.63	212	0.080
c_26	52.21	7.68	181	0.581
c_27	30.46	32.60	234	0.280
c_28	11.91	78.23	242	0.462
c_29	-8.19	-35.31	142	0.060
c_30	45.80	-122.36	180	0.097
c_31	-33.61	-70.97	60	0.208
c_32	29.34	47.95	60	0.327
c_33	42.60	141.79	134	0.352
c_34	13.51	100.38	72	0.462
c_35	6.51	3.08	121	0.205
c_36	-15.77	-48.59	181	0.042
c_37	38.26	-121.97	68	0.037
c_38	39.97	-90.59	183	0.044
c_39	24.43	55.51	206	0.148
c_40	-1.12	36.23	136	0.233
c_41	40.83	14.20	181	0.585
c_42	-25.64	28.25	82	0.119
c_43	9.78	8.48	189	0.258
c_44	39.39	-105.17	72	0.077
c_45	18.39	-68.93	217	0.162
c_46	-2.20	-78.77	212	0.272
c_47	21.75	38.97	79	0.132
c_48	-32.40	152.52	541	0.138
c_49	32.36	-110.91	147	0.034
c_50	14.57	-89.13	201	0.073
c_51	23.29	88.22	219	0.341
c_52	16.22	103.14	323	0.132
c_53	7.59	-4.06	411	0.240
c_54	21.04	-103.48	121	0.034
c_55	44.63	20.19	246	0.651
c_56	51.14	19.08	241	0.670
c_57	9.70	123.98	147	0.324
c_58	38.59	-9.32	124	0.243
c_59	30.47	30.08	183	0.360
c_60	-1.12	-48.36	60	0.071
c_61	45.36	-93.33	121	0.090
c_62	-28.37	24.25	61	0.099
c_63	-20.66	-48.68	179	0.012
c_64	6.46	125.29	125	0.075
c_65	42.97	28.21	243	0.338
c_66	-8.69	114.90	121	0.204
c_67	-25.52	-57.66	69	0.139
c_68	8.63	-79.79	121	0.144
c_69	24.84	122.00	266	0.528
c_70	10.26	-74.73	124	0.162
c_71	-7.61	112.44	77	0.367
c_72	-40.10	-70.88	317	0.065
c_73	-28.75	-64.77	302	0.107
c_74	10.26	-67.57	121	0.205
c_75	-37.81	144.80	60	0.158
c_76	11.90	-61.94	307	0.033
c_77	25.91	-100.14	60	0.131
c_78	-13.07	-38.36	62	0.145
c_79	47.30	-0.34	199	0.481
c_80	-29.99	30.64	184	0.305
c_81	5.42	8.00	219	0.116
c_82	20.59	82.80	167	0.283
c_83	36.98	35.29	145	0.262
c_84	-0.03	115.38	387	0.195
c_85	37.18	-122.01	0	0.123
c_86	39.67	41.67	212	0.296
c_87	-22.82	-52.60	219	0.153
c_88	-2.75	-58.73	60	0.046
c_89	53.74	49.52	784	0.601
c_90	-16.38	31.45	422	0.219
c_91	37.26	126.75	61	0.833
c_92	-1.67	29.19	459	0.350
c_93	-3.84	-38.17	83	0.125
c_94	40.53	116.28	1131	0.636
c_95	-3.29	-43.56	139	0.070
c_96	33.62	72.91	166	0.142
c_97	23.30	78.19	112	0.286
c_98	17.34	78.66	172	0.376
c_99	20.52	-88.01	155	0.128
c_100	48.52	-114.05	391	0.189
c_101	42.43	2.42	79	0.458
c_102	49.03	-122.88	0	0.142
c_103	59.81	18.85	194	0.616
c_104	25.50	67.87	104	0.483
c_105	6.47	100.20	128	0.127
c_106	-17.96	-40.22	181	0.082
c_107	41.57	-96.11	74	0.132
c_108	-6.57	39.09	61	0.547
c_109	-22.38	23.83	181	0.222
c_110	-10.33	-64.28	411	0.040
c_111	-11.95	-75.81	60	0.196
c_112	40.44	-112.08	66	0.091
c_113	18.94	97.02	121	0.605
c_114	44.36	-68.61	252	0.078
c_115	18.38	-76.93	67	0.163
c_116	41.32	36.54	147	0.024
c_117	-8.18	-48.45	312	0.078
c_118	55.95	-4.68	148	0.259
c_119	56.02	12.72	168	0.389
c_120	39.17	3.05	61	0.188
c_121	17.93	43.54	203	0.318
c_122	37.01	23.09	120	0.729