deberta_MP_dynamic by Oriane Peter: Open-Weight Model

Runs On

What it takes to serve deberta_MP_dynamic (435M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	0.9 GB	1.0 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
8-bit	0.4 GB	0.5 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
4-bit	0.2 GB	0.3 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 1, 2026.

deberta_MP_dynamic on every accelerator the SAVRN Index prices, at every precision

Model Card

By Oriane Peter, published under mit, revision 31d3e69f9dc9.

This model is a fine-tuned version of microsoft/deberta-v3-large on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 3e-05 - trainbatchsize: 16 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 32 - lrschedulertype: cosine - lrschedulerwarmupsteps: 0.1 - numepochs: 1000 - Transformers 5.12.1 - Pytorch 2.11.0+cu128 - Datasets 5.0.1 - Tokenizers 0.22.2

Read Oriane Peter's full model card

This model is a fine-tuned version of microsoft/deberta-v3-large on an unknown dataset. It achieves the following results on the evaluation set: - Loss: 0.1224 - Macro F1: 0.5993 - Micro F1: 0.6698 - Macro Precision: 0.5833 - Macro Recall: 0.6198 - Micro Precision: 0.6508 - Micro Recall: 0.6898 - Exact Match Ratio: 0.0756 - Macro Roc Auc: 0.8980

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training: - learning_rate: 3e-05 - train_batch_size: 16 - eval_batch_size: 16 - seed: 42 - gradient_accumulation_steps: 2 - total_train_batch_size: 32 - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments - lr_scheduler_type: cosine - lr_scheduler_warmup_steps: 0.1 - num_epochs: 1000

Training results

Training Loss	Epoch	Step	Validation Loss	Macro F1	Micro F1	Macro Precision	Macro Recall	Micro Precision	Micro Recall	Exact Match Ratio	Macro Roc Auc
0.1465	12.8312	500	0.1330	0.0747	0.2343	0.1856	0.0569	0.5726	0.1473	0.0091	0.7400
0.1375	25.6494	1000	0.1276	0.3216	0.5160	0.4262	0.3049	0.5651	0.4747	0.0287	0.8359
0.1341	38.4675	1500	0.1264	0.3569	0.5434	0.5077	0.3403	0.6199	0.4838	0.0415	0.8623
0.1311	51.2857	2000	0.1256	0.3895	0.5649	0.5364	0.3516	0.6622	0.4926	0.0550	0.8777
0.1292	64.1039	2500	0.1239	0.4202	0.5950	0.5565	0.3972	0.6184	0.5733	0.0403	0.8882
0.1282	76.9351	3000	0.1230	0.4595	0.6156	0.5226	0.4707	0.5995	0.6326	0.0375	0.8948
0.1261	89.7532	3500	0.1226	0.4813	0.6258	0.5394	0.4899	0.6158	0.6362	0.0438	0.8991
0.1244	102.5714	4000	0.1224	0.4999	0.6321	0.5574	0.4918	0.6251	0.6392	0.0513	0.8987
0.1236	115.3896	4500	0.1228	0.5070	0.6272	0.5033	0.5565	0.5710	0.6956	0.0293	0.9031
0.1226	128.2078	5000	0.1218	0.5264	0.6449	0.5336	0.5453	0.6114	0.6822	0.0427	0.9045
0.1212	141.0260	5500	0.1216	0.5281	0.6490	0.5699	0.5257	0.6485	0.6495	0.0504	0.9062
0.1199	153.8571	6000	0.1213	0.5410	0.6510	0.5528	0.5639	0.6220	0.6827	0.0507	0.9078
0.1189	166.6753	6500	0.1216	0.5464	0.6513	0.5374	0.5807	0.6130	0.6946	0.0463	0.9085
0.1183	179.4935	7000	0.1218	0.5480	0.6529	0.5846	0.5361	0.6646	0.6415	0.0559	0.9012
0.1180	192.3117	7500	0.1214	0.5578	0.6552	0.5453	0.5893	0.6191	0.6957	0.0453	0.9069
0.1170	205.1299	8000	0.1214	0.5594	0.6579	0.5841	0.5593	0.6608	0.6550	0.0594	0.9067
0.1163	217.9610	8500	0.1214	0.5698	0.6573	0.5460	0.6099	0.6075	0.7158	0.0444	0.9045
0.1151	230.7792	9000	0.1214	0.5733	0.6621	0.5715	0.5900	0.6491	0.6757	0.0559	0.9061
0.1150	243.5974	9500	0.1216	0.5749	0.6611	0.5524	0.6126	0.6266	0.6996	0.0494	0.9059
0.1147	256.4156	10000	0.1214	0.5796	0.6620	0.5694	0.6046	0.6354	0.6909	0.0572	0.9064
0.1141	269.2338	10500	0.1216	0.5753	0.6614	0.5817	0.5872	0.6524	0.6706	0.0590	0.9000
0.1137	282.0519	11000	0.1212	0.5830	0.6647	0.5830	0.5950	0.6505	0.6796	0.0628	0.9039
0.1133	294.8831	11500	0.1215	0.5792	0.6635	0.5742	0.5980	0.6497	0.6779	0.0600	0.9028
0.1130	307.7013	12000	0.1217	0.5839	0.6641	0.5602	0.6250	0.6283	0.7042	0.0545	0.9048
0.1131	320.5195	12500	0.1216	0.5812	0.6624	0.5705	0.6051	0.6436	0.6824	0.0548	0.9040
0.1121	333.3377	13000	0.1217	0.5875	0.6628	0.5576	0.6324	0.6254	0.7049	0.0516	0.9044
0.1119	346.1558	13500	0.1217	0.5869	0.6642	0.5654	0.6192	0.6437	0.6862	0.0653	0.9053
0.1116	358.9870	14000	0.1219	0.5896	0.6648	0.5710	0.6203	0.6361	0.6963	0.0568	0.9043
0.1113	371.8052	14500	0.1217	0.5900	0.6669	0.5709	0.6227	0.6419	0.6938	0.0608	0.9024
0.1109	384.6234	15000	0.1216	0.5922	0.6677	0.5722	0.6226	0.6498	0.6865	0.0607	0.9053
0.1105	397.4416	15500	0.1219	0.5900	0.6656	0.5700	0.6243	0.6400	0.6934	0.0593	0.9052
0.1103	410.2597	16000	0.1219	0.5899	0.6670	0.5743	0.6186	0.6447	0.6909	0.0616	0.9021
0.1099	423.0779	16500	0.1215	0.5927	0.6700	0.5942	0.5990	0.6696	0.6704	0.0694	0.8993
0.1098	435.9091	17000	0.1220	0.5873	0.6662	0.5714	0.6160	0.6501	0.6831	0.0662	0.9013
0.1093	448.7273	17500	0.1220	0.5902	0.6681	0.5687	0.6213	0.6435	0.6947	0.0634	0.8999
0.1094	461.5455	18000	0.1219	0.5931	0.6681	0.5787	0.6160	0.6477	0.6899	0.0650	0.8996
0.1092	474.3636	18500	0.1217	0.5928	0.6703	0.5956	0.5996	0.6690	0.6717	0.0700	0.8976
0.1088	487.1818	19000	0.1222	0.5919	0.6675	0.5822	0.6128	0.6547	0.6808	0.0648	0.8988
0.1086	500.0	19500	0.1225	0.5897	0.6652	0.5699	0.6200	0.6505	0.6806	0.0695	0.9002
0.1082	512.8312	20000	0.1219	0.5950	0.6692	0.5819	0.6169	0.6535	0.6857	0.0691	0.9007
0.1081	525.6494	20500	0.1217	0.5937	0.6695	0.5875	0.6110	0.6609	0.6783	0.0692	0.8997
0.1082	538.4675	21000	0.1224	0.5973	0.6683	0.5683	0.6364	0.6373	0.7026	0.0666	0.9016
0.1082	551.2857	21500	0.1220	0.5962	0.6691	0.5803	0.6212	0.6526	0.6865	0.0663	0.9011
0.1075	564.1039	22000	0.1222	0.5995	0.6695	0.5884	0.6186	0.6631	0.6760	0.0701	0.8978
0.1075	576.9351	22500	0.1221	0.5989	0.6705	0.5825	0.6231	0.6495	0.6928	0.0664	0.9000
0.1074	589.7532	23000	0.1219	0.5994	0.6734	0.5798	0.6260	0.6519	0.6965	0.0683	0.8998
0.1075	602.5714	23500	0.1220	0.6010	0.6721	0.5843	0.6236	0.6521	0.6934	0.0668	0.8953
0.1074	615.3896	24000	0.1222	0.5996	0.6721	0.5816	0.6256	0.6503	0.6954	0.0686	0.8985
0.1069	628.2078	24500	0.1224	0.5963	0.6696	0.5810	0.6200	0.6582	0.6814	0.0696	0.8981
0.1068	641.0260	25000	0.1224	0.5983	0.6710	0.5869	0.6172	0.6603	0.6820	0.0712	0.8958
0.1068	653.8571	25500	0.1221	0.6029	0.6723	0.5975	0.6141	0.6668	0.6778	0.0753	0.8958
0.1065	666.6753	26000	0.1221	0.5980	0.6700	0.6006	0.6033	0.6730	0.6670	0.0795	0.8947
0.1065	679.4935	26500	0.1224	0.6024	0.6728	0.5847	0.6273	0.6523	0.6947	0.0707	0.8967
0.1064	692.3117	27000	0.1223	0.6027	0.6739	0.5805	0.6318	0.6530	0.6962	0.0707	0.8977
0.1063	705.1299	27500	0.1225	0.6006	0.6704	0.5936	0.6144	0.6681	0.6726	0.0765	0.8947
0.1060	717.9610	28000	0.1226	0.6036	0.6723	0.5818	0.6310	0.6512	0.6948	0.0721	0.8969
0.1062	730.7792	28500	0.1223	0.6031	0.6724	0.5838	0.6286	0.6533	0.6926	0.0723	0.8956
0.1058	743.5974	29000	0.1223	0.6033	0.6728	0.5984	0.6122	0.6699	0.6758	0.0776	0.8939
0.1059	756.4156	29500	0.1224	0.6055	0.6738	0.5971	0.6217	0.6621	0.6859	0.0768	0.8935
0.1056	769.2338	30000	0.1224	0.6038	0.6736	0.5961	0.6165	0.6659	0.6814	0.0768	0.8936
0.1056	782.0519	30500	0.1224	0.6036	0.6731	0.5867	0.6252	0.6576	0.6892	0.0731	0.8957
0.1056	794.8831	31000	0.1225	0.6060	0.6745	0.5837	0.6345	0.6527	0.6979	0.0717	0.8978
0.1055	807.7013	31500	0.1225	0.6026	0.6737	0.5822	0.6289	0.6552	0.6934	0.0737	0.8969
0.1054	820.5195	32000	0.1225	0.6037	0.6735	0.5856	0.6278	0.6554	0.6927	0.0736	0.8964
0.1054	833.3377	32500	0.1225	0.6046	0.6747	0.5853	0.6295	0.6550	0.6957	0.0762	0.8959
0.1054	846.1558	33000	0.1226	0.6034	0.6737	0.5841	0.6295	0.6560	0.6924	0.0751	0.8962
0.1052	858.9870	33500	0.1224	0.6057	0.6740	0.5944	0.6228	0.6662	0.6819	0.0773	0.8948
0.1050	871.8052	34000	0.1226	0.6042	0.6744	0.5891	0.6253	0.6607	0.6886	0.0756	0.8954
0.1053	884.6234	34500	0.1226	0.6052	0.6752	0.5878	0.6287	0.6573	0.6942	0.0758	0.8955
0.1052	897.4416	35000	0.1226	0.6050	0.6754	0.5887	0.6264	0.6600	0.6916	0.0756	0.8953
0.1052	910.2597	35500	0.1226	0.6049	0.6745	0.5862	0.6292	0.6586	0.6912	0.0760	0.8963
0.1051	923.0779	36000	0.1225	0.6061	0.6754	0.5904	0.6268	0.6616	0.6898	0.0772	0.8954
0.1052	935.9091	36500	0.1226	0.6050	0.6751	0.5882	0.6272	0.6596	0.6914	0.0755	0.8960
0.1050	948.7273	37000	0.1226	0.6054	0.6752	0.5867	0.6301	0.6585	0.6927	0.0750	0.8960
0.1052	961.5455	37500	0.1225	0.6060	0.6753	0.5897	0.6275	0.6602	0.6912	0.0753	0.8959
0.1050	974.3636	38000	0.1225	0.6065	0.6754	0.5887	0.6296	0.6595	0.6920	0.0759	0.8961
0.1053	987.1818	38500	0.1225	0.6057	0.6752	0.5884	0.6284	0.6591	0.6921	0.0760	0.8961
0.1051	1000.0	39000	0.1225	0.6060	0.6754	0.5888	0.6286	0.6595	0.6921	0.0759	0.8961

Framework versions

Transformers 5.12.1
Pytorch 2.11.0+cu128
Datasets 5.0.1
Tokenizers 0.22.2

Configuration

Architecture: DebertaV2ForSequenceClassification
Context length (tokens): 512
Layers: 24
Hidden size: 1,024
Feed-forward size: 4,096
Attention heads: 16
Vocabulary size: 128,100
Model type: deberta-v2

Identity and Version

Repository: orpe42/deberta_MP_dynamic
Publisher: Oriane Peter
Task: Text classification
Modality: Text
Library: transformers
Parameters: 435M parameters
Languages: Not stated by the source
Revision: 31d3e69f9dc99210bae75efdfb941a199cf8a554
First published: 2026-08-06
Last updated: 2026-09-30

Files and Weights

44 files, 8.7 GB in total. The weights are 9 files totalling 8.7 GB in bin, pt, pth, safetensors.

Weights9 files · 8.7 GB

Configuration4 files · 132.2 KB

Tokenizer6 files · 25.0 MB

Documentation1 file · 15.9 KB

Other23 files · 1.2 MB

Repository1 file · 1.5 KB

Every file

File	Type	Size	SHA-256
best_model/model.safetensors	Weights	1.7 GB	190423dea1d6
best_model/training_args.bin	Weights	5.3 KB	a4b7760eded9
current_best/model.safetensors	Weights	1.7 GB	bc6248cc5ff2
current_best/optimizer.pt	Weights	3.5 GB	098933cebcc7
current_best/rng_state.pth	Weights	14.2 KB	96a3667d6833
current_best/scheduler.pt	Weights	1.1 KB	7eb7ccbcbe37
current_best/training_args.bin	Weights	4.9 KB	4c5264262098
model.safetensors	Weights	1.7 GB	5a80c69110e6
training_args.bin	Weights	4.9 KB	ebeedcee40d8
best_model/config.json	Configuration	5.1 KB	—
config.json	Configuration	3.6 KB	—
current_best/config.json	Configuration	5.1 KB	—
current_best/trainer_state.json	Configuration	118.2 KB	—
README.md	Documentation	15.9 KB	—
best_thresholds.npy	Other	576 B	7b874082e7c2
runs/Aug06_12-10-47_erc-hpc-vm042/events.out.tfevents.1786014647.erc-hpc-vm042.307913.0	Other	63.3 KB	3c1474c7be85
runs/Aug06_12-10-47_erc-hpc-vm042/events.out.tfevents.1786031039.erc-hpc-vm042.307913.1	Other	811 B	65be6cb94ebc
runs/Aug06_12-13-56_erc-hpc-vm044/events.out.tfevents.1786014836.erc-hpc-vm044.79796.0	Other	21.9 KB	d3640ea97f1d
runs/Aug06_14-27-22_erc-hpc-comp031/events.out.tfevents.1786022842.erc-hpc-comp031.2297350.0	Other	8.8 KB	bbc23ae94833
runs/Aug06_15-42-04_erc-hpc-comp031/events.out.tfevents.1786027324.erc-hpc-comp031.2302155.0	Other	49.5 KB	edca708fd60a
runs/Aug06_15-42-04_erc-hpc-comp031/events.out.tfevents.1786050127.erc-hpc-comp031.2302155.1	Other	811 B	d9f33ae0fe41
runs/Aug06_16-58-04_erc-hpc-comp248/events.out.tfevents.1786031884.erc-hpc-comp248.1390536.0	Other	90.9 KB	020d926d2d43
runs/Aug06_16-58-04_erc-hpc-comp248/events.out.tfevents.1786073374.erc-hpc-comp248.1390536.1	Other	824 B	ea33405baff6
runs/Aug06_23-42-40_erc-hpc-vm041/events.out.tfevents.1786056160.erc-hpc-vm041.4006523.0	Other	67.5 KB	39a2ec7ae027
runs/Aug07_08-40-50_erc-hpc-vm041/events.out.tfevents.1786088450.erc-hpc-vm041.4016584.0	Other	14.1 KB	1b7ba8284dd5
runs/Aug07_08-44-07_erc-hpc-comp248/events.out.tfevents.1786088647.erc-hpc-comp248.1584800.0	Other	12.3 KB	b04d9f9bbc76
runs/Aug07_09-31-05_erc-hpc-vm041/events.out.tfevents.1786091465.erc-hpc-vm041.4017016.0	Other	31.9 KB	f262cdd89e0e
runs/Aug07_09-32-07_erc-hpc-comp248/events.out.tfevents.1786091527.erc-hpc-comp248.1586555.0	Other	30.1 KB	c6cbe1e8c47d
runs/Aug18_15-19-24_erc-hpc-comp248/events.out.tfevents.1787062764.erc-hpc-comp248.1575213.0	Other	145.4 KB	dc621e4ea00b
runs/Aug19_14-33-28_erc-hpc-comp033/events.out.tfevents.1787146408.erc-hpc-comp033.100299.0	Other	149.3 KB	6209a821ec74
runs/Aug19_14-33-28_erc-hpc-comp033/events.out.tfevents.1787222161.erc-hpc-comp033.100299.1	Other	824 B	e5929b9e8b37
runs/Aug23_16-46-45_erc-hpc-vm042/events.out.tfevents.1787500005.erc-hpc-vm042.2697604.0	Other	124.0 KB	523c970927f1
runs/Aug24_08-29-23_erc-hpc-comp031/events.out.tfevents.1787556563.erc-hpc-comp031.186562.0	Other	149.3 KB	ff18ab3068d4
runs/Aug24_08-29-23_erc-hpc-comp031/events.out.tfevents.1787633956.erc-hpc-comp031.186562.1	Other	824 B	097ba003dba0
runs/Sep26_14-51-43_erc-hpc-comp246/events.out.tfevents.1790430703.erc-hpc-comp246.28815.0	Other	149.0 KB	1341975c1a80
runs/Sep26_14-51-43_erc-hpc-comp246/events.out.tfevents.1790453924.erc-hpc-comp246.28815.1	Other	824 B	7ab39a74dbf7
runs/Sep29_13-20-28_erc-hpc-comp039/events.out.tfevents.1790684428.erc-hpc-comp039.1459476.0	Other	93.1 KB	74cccc7916f7
.gitattributes	Repository	1.5 KB	—
best_model/tokenizer.json	Tokenizer	8.3 MB	—
best_model/tokenizer_config.json	Tokenizer	538 B	—
current_best/tokenizer.json	Tokenizer	8.3 MB	—
current_best/tokenizer_config.json	Tokenizer	538 B	—
tokenizer.json	Tokenizer	8.3 MB	—
tokenizer_config.json	Tokenizer	538 B	—

License and Download

License: mit
Access: Open weights, no gate
Download size: 8.7 GB

Download from Oriane Peter

Released by Oriane Peter through its official repository on Hugging Face. Read the license.

Built From

Derived from microsoft/deberta-v3-large

Memory Requirements

Precision	Weights in memory
As published	8.7 GB
16-bit	0.9 GB
8-bit	0.4 GB
4-bit	0.2 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About deberta_MP_dynamic

How much GPU memory does deberta_MP_dynamic need?

About 1 GB at 16-bit and 0.3 GB at 4-bit: the weights (435M parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run deberta_MP_dynamic on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use deberta_MP_dynamic commercially?

Yes. deberta_MP_dynamic is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

What is deberta_MP_dynamic's context length?

512 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Text classification

DEBATE-kor-large

Jong Rock Jeong

DEBATE-kor-large is a Korean-adapted Political DEBATE model for binary natural language inference (NLI) on political text. The model is initialized from mlburnham/PoliticalDEBATEDeBERTalargev1.1, the original DeBERTa-based Political DEBATE checkpoint, and subsequently fine-tuned on jongrock17/PolNLI-kor, a Korean translation and adaptation of PolNLI. Political DEBATE DeBERTa-large → PolNLI-kor → DEBATE-kor-large Unlike the PolNLI-kor-RoBERTa model family, which starts from Korean-pretrained KLUE-RoBERTa encoders, DEBATE-kor directly adapts the original Political DEBATE checkpoint to Korean political NLI. DEBATE-kor formulates NLI as a binary classification problem. notentailment combines…

Open weights 435M parameters 512 tokens transformers

View model

Model · Text classification

opir-multitask-large-v1.0

Knowledgator Engineering

Opir-multitask-large is the English, highest-accuracy multi-task checkpoint in the Opir family: an encoder-based GLiClass guardrail model for real-time LLM safety filtering. It supports binary safe/unsafe classification, toxicity detection, jailbreak and prompt-injection detection, and zero-shot harmful-content categorization over a hierarchical safety taxonomy. This card is for knowledgator/opir-multitask-large. The model is used through GLiClass zero-shot classification: pass text plus the candidate labels you want scored. Use single-label mode for binary safe/unsafe decisions and multi-label mode for taxonomy, toxicity, jailbreak, or custom policy labels. Use multi-label mode when you…

Open weights apache-2.0 439M parameters gliclass

View model

Model · Text classification

erabi-practical-v1-experimental

Sugarknight

This is an experimental, uncalibrated choice-ranking model. It is not an official Jev model, a validated general-purpose reasoner, or an automatic decision-maker. The model ranks 2–16 user-supplied candidate texts for a natural-language context and question and returns all candidate probabilities through the ERABI code. Decisions should be reviewed by a person. - Practical V1 data consists of original synthetic Japanese, English, and Simplified Chinese examples in six task families, generated and answer-blind rejudged with DeepSeek V4.1 Flash. - The Exam-QA source was filtered and transformed with the same DeepSeek model. Symbolic answer labels were mapped to source choice text. Ambiguous…

Open weights apache-2.0 439M parameters transformers

View model

Model · Text classification

laya-typed-decisions

Convai Innovations

Non-autoregressive System 1 decision model, fine-tuned on the typed-decisions workflows: agent-trace observability, customer service, invoice processing and security incidents. Part of the Laya family. 400 test cases, 2,000 decisions, measured on the official test split. +3.9 points over Jev's published 0.727, above the 0.735 teacher ceiling, with 2.4x better Brier and 1.6x better score MAE. Jev figures are third-party published, not measured here — there is no TypeSafe API access in this project, and sample sizes and prompts differ. Treat the comparison as indicative. Router will not select this checkpoint automatically unless you construct it with autotaskdetection=True — it is…

Open weights apache-2.0 421M parameters transformers

View model

Model · Text classification

laya

Convai Innovations

Multilingual, non-autoregressive System 1 decision model. Give it a state (text, email, ticket, or JSON) and typed questions; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages. Trained with reinforcement learning against strictly proper scoring rules (RLCD), so reporting honest probabilities is the only way to maximise reward. It never generates text, so there is nothing to parse and nothing to hallucinate. This repo holds all three checkpoints and is the hub for the family. The English checkpoint is at the repo root; the other two are bundled subfolders, and only the one you request is downloaded: pip install -U…

Open weights apache-2.0 421M parameters transformers

View model

Model · Text classification

laya-kvp10k-noul

Sothiara Em

A fine-tuned Laya model (Convai Innovations, 421M, ModernBERT-large backbone) that answers one typed noul question: does a value correctly match its key label? (e.g. firstname = John → true, firstname = 1992 → false). Fine-tuned on the IBM KVP-10K dataset via the pre-parsed community mirror (OCR pre-extracted; no OCR step needed). This is the v2 (remediated) release: document-level 80/10/10 data partitioning, mixed-class frozen test set, calib-split temperature fitting, and a frozen category-stratified release gate. It supersedes the v1 release pre-remediation, historical artifact). Laya is a non-autoregressive decision model: you give it a state (text/dict) plus typed questions (choice…

Open weights apache-2.0 421M parameters laya

View model