Multilingual Verdict Classifier

This model is a fine-tuned version of xlm-roberta-base on 2,500 deduplicated multilingual verdicts from Google Fact Check Tools API, translated into 65 languages with the Google Cloud Translation API. It achieves the following results on the evaluation set, being 1,000 such verdicts, but here including duplicates to represent the true distribution:

Loss: 0.2238
F1 Macro: 0.8540
F1 Misinformation: 0.9798
F1 Factual: 0.9889
F1 Other: 0.5934
Prec Macro: 0.8348
Prec Misinformation: 0.9860
Prec Factual: 0.9889
Prec Other: 0.5294

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 162525
num_epochs: 1000

Training results

Training Loss	Epoch	Step	Validation Loss	F1 Macro	F1 Misinformation	F1 Factual	F1 Other	Prec Macro	Prec Misinformation	Prec Factual	Prec Other
1.1109	0.1	2000	1.2166	0.0713	0.1497	0.0	0.0640	0.2451	0.7019	0.0	0.0334
0.9551	0.2	4000	0.7801	0.3611	0.8889	0.0	0.1943	0.3391	0.8915	0.0	0.1259
0.9275	0.3	6000	0.7712	0.3468	0.9123	0.0	0.1282	0.3304	0.9051	0.0	0.0862
0.8881	0.39	8000	0.5386	0.3940	0.9524	0.0	0.2297	0.3723	0.9748	0.0	0.1420
0.7851	0.49	10000	0.3298	0.6886	0.9626	0.7640	0.3393	0.6721	0.9798	0.7727	0.2639
0.639	0.59	12000	0.2156	0.7847	0.9633	0.9355	0.4554	0.7540	0.9787	0.9062	0.3770
0.5677	0.69	14000	0.1682	0.7877	0.9694	0.9667	0.4270	0.7763	0.9745	0.9667	0.3878
0.5218	0.79	16000	0.1475	0.8037	0.9692	0.9667	0.4752	0.7804	0.9812	0.9667	0.3934
0.4682	0.89	18000	0.1458	0.8097	0.9734	0.9667	0.4889	0.7953	0.9791	0.9667	0.44
0.4188	0.98	20000	0.1416	0.8370	0.9769	0.9724	0.5618	0.8199	0.9826	0.9670	0.5102
0.3735	1.08	22000	0.1624	0.8094	0.9698	0.9368	0.5217	0.7780	0.9823	0.89	0.4615
0.3242	1.18	24000	0.1648	0.8338	0.9769	0.9727	0.5517	0.8167	0.9826	0.9570	0.5106
0.2785	1.28	26000	0.1843	0.8261	0.9739	0.9780	0.5263	0.8018	0.9836	0.9674	0.4545
0.25	1.38	28000	0.1975	0.8344	0.9744	0.9834	0.5455	0.8072	0.9859	0.9780	0.4576
0.2176	1.48	30000	0.1849	0.8209	0.9691	0.9889	0.5047	0.7922	0.9846	0.9889	0.4030
0.1966	1.58	32000	0.2119	0.8194	0.9685	0.9944	0.4954	0.7920	0.9846	1.0	0.3913
0.1738	1.67	34000	0.2110	0.8352	0.9708	0.9944	0.5405	0.8035	0.9881	1.0	0.4225
0.1625	1.77	36000	0.2152	0.8165	0.9709	0.9834	0.4950	0.7905	0.9835	0.9780	0.4098
0.1522	1.87	38000	0.2300	0.8097	0.9697	0.9832	0.4762	0.7856	0.9835	0.9888	0.3846
0.145	1.97	40000	0.1955	0.8519	0.9774	0.9889	0.5895	0.8280	0.9860	0.9889	0.5091
0.1248	2.07	42000	0.2308	0.8149	0.9703	0.9889	0.4854	0.7897	0.9835	0.9889	0.3968
0.1186	2.17	44000	0.2368	0.8172	0.9733	0.9834	0.4948	0.7942	0.9836	0.9780	0.4211
0.1122	2.26	46000	0.2401	0.7968	0.9804	0.8957	0.5143	0.8001	0.9849	1.0	0.4154
0.1099	2.36	48000	0.2290	0.8119	0.9647	0.9834	0.4874	0.7777	0.9880	0.9780	0.3671
0.1093	2.46	50000	0.2256	0.8247	0.9745	0.9889	0.5106	0.8053	0.9825	0.9889	0.4444
0.1053	2.56	52000	0.2416	0.8456	0.9799	0.9889	0.5679	0.8434	0.9805	0.9889	0.5610
0.1049	2.66	54000	0.2850	0.7585	0.9740	0.8902	0.4112	0.7650	0.9802	0.9865	0.3284
0.098	2.76	56000	0.2828	0.8049	0.9642	0.9889	0.4615	0.7750	0.9856	0.9889	0.3506
0.0962	2.86	58000	0.2238	0.8540	0.9798	0.9889	0.5934	0.8348	0.9860	0.9889	0.5294
0.0975	2.95	60000	0.2494	0.8249	0.9715	0.9889	0.5143	0.7967	0.9858	0.9889	0.4154
0.0877	3.05	62000	0.2464	0.8274	0.9733	0.9889	0.5200	0.8023	0.9847	0.9889	0.4333
0.0848	3.15	64000	0.2338	0.8263	0.9740	0.9889	0.5161	0.8077	0.9814	0.9889	0.4528
0.0859	3.25	66000	0.2335	0.8365	0.9750	0.9889	0.5455	0.8108	0.9859	0.9889	0.4576
0.084	3.35	68000	0.2067	0.8343	0.9763	0.9889	0.5376	0.8148	0.9837	0.9889	0.4717
0.0837	3.45	70000	0.2516	0.8249	0.9746	0.9889	0.5111	0.8097	0.9803	0.9889	0.46
0.0809	3.54	72000	0.2948	0.8258	0.9728	0.9944	0.5102	0.8045	0.9824	1.0	0.4310
0.0833	3.64	74000	0.2457	0.8494	0.9744	0.9944	0.5794	0.8173	0.9893	1.0	0.4627
0.0796	3.74	76000	0.3188	0.8277	0.9733	0.9889	0.5208	0.8059	0.9825	0.9889	0.4464
0.0821	3.84	78000	0.2642	0.8343	0.9714	0.9944	0.5370	0.8045	0.9870	1.0	0.4265

Framework versions

Transformers 4.11.3
Pytorch 1.9.0+cu102
Datasets 1.9.0
Tokenizers 0.10.2

saattrupdan/verdict-classifier

Multilingual Verdict Classifier

Training procedure

Training hyperparameters

Training results

Framework versions

Seed this model