dense_reward_trainer_final_opt__NumTrainEpochs5_SaveStrategiesepoch_reward_modeling_anthropic_hh

cj453/dense_reward_trainer_final_opt__NumTrainEpochs5_SaveStrategiesepoch_reward_modeling_anthropic_hh

dense_reward_trainer_final_opt__NumTrainEpochs5_SaveStrategiesepoch_reward_modeling_anthropic_hh is a machine learning model from cj453. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.

should probably proofread and complete it, then remove this comment. --

Downloads · 30 days

12% of all-time downloads

All-time downloads

337

Public

Parameters

1.4B

5.7 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors5.7 GB · 100%

At a glance

License: other
Model type: opt
Access: Public
Created: Sep 16, 2024
Updated: Sep 16, 2024
SHA: 14b66987

Base models

facebook/opt-1.3b

Model card

From the Hugging Face model README

dense_reward_trainer_final_opt__NumTrainEpochs5_SaveStrategiesepoch_reward_modeling_anthropic_hh

This model is a fine-tuned version of facebook/opt-1.3b on an unknown dataset. It achieves the following results on the evaluation set:

Loss: 2.4124
Accuracy: 0.6660
Train Rewards/chosen: 9.2061
Train Rewards/rejected: -9.4536
Train Rewards/accuracies: 0.9844
Train Rewards/margins: 18.6597
Train Nll Loss: 2.1547
Train Logit Total Loss: 0.0587
Train Logit Loss: 0.0375
Rewards/chosen: 3.4303
Rewards/rejected: -2.3575
Rewards/accuracies: 0.6484
Rewards/margins: 5.7878
Nll Loss: 2.1950
Logit Total Loss: 2.4421
Logit Loss: 2.4446

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: 1.41e-05
train_batch_size: 4
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Nll Loss	Logit Total Loss	Logit Loss
0.7077	0.11	100	0.6897	0.6165	-1.6378	-1.8001	0.6016	0.1622	2.8092	0.6881	0.6667
0.7117	0.23	200	0.6764	0.6103	-2.8148	-3.0536	0.5964	0.2388	2.8927	0.6746	0.6522
0.6502	0.34	300	0.6626	0.6536	-0.8018	-1.1645	0.6399	0.3627	2.9696	0.6611	0.6377
0.655	0.46	400	0.6503	0.6144	-1.5457	-1.9648	0.5984	0.4191	2.7773	0.6489	0.6274
0.6467	0.57	500	0.6653	0.6165	-0.9541	-1.3483	0.6036	0.3942	2.8139	0.6643	0.6426
0.6694	0.69	600	0.6432	0.6392	-1.5917	-1.9439	0.6278	0.3522	2.7779	0.6426	0.6211
0.6753	0.8	700	0.6494	0.6371	-1.3508	-1.7191	0.6246	0.3683	2.8056	0.6474	0.6256
0.6806	0.91	800	0.6449	0.6103	-1.4576	-1.8165	0.6004	0.3589	2.7215	0.6424	0.6214
0.5434	1.03	900	0.6827	0.6557	-0.8965	-1.6611	0.6468	0.7645	2.6762	0.6816	0.6615
0.5448	1.14	1000	0.7194	0.6392	-0.8661	-1.8265	0.6266	0.9604	2.6214	0.7184	0.6992
0.5129	1.26	1100	0.7990	0.6289	1.3108	0.2390	0.6165	1.0718	2.6526	0.7966	0.7779
0.5033	1.37	1200	0.6888	0.6557	-0.9571	-1.8601	0.6488	0.9030	2.6263	0.6868	0.6672
0.404	1.49	1300	0.7422	0.6309	-1.1408	-2.0297	0.6226	0.8890	2.6046	0.7348	0.7159
0.5512	1.6	1400	0.6762	0.6474	-2.5166	-3.3023	0.6327	0.7857	2.5872	0.6766	0.6573
0.4558	1.71	1500	0.6843	0.6619	-2.3183	-3.2412	0.6476	0.9229	2.5268	0.6811	0.6625
0.5184	1.83	1600	0.7135	0.6557	-1.5991	-2.5538	0.6456	0.9547	2.5671	0.7179	0.6992
0.4213	1.94	1700	0.7220	0.6495	-1.3947	-2.4198	0.6395	1.0251	2.5040	0.7198	0.7018
0.1508	2.06	1800	1.0827	0.6598	2.5282	0.2534	0.6476	2.2748	2.6437	1.0758	1.0599
0.1216	2.17	1900	1.1376	0.6474	-0.0750	-2.1523	0.6302	2.0773	2.5506	1.1502	1.1361
0.1044	2.29	2000	1.4682	0.6722	-0.4860	-3.5268	0.6577	3.0408	2.5292	1.4836	1.4730
0.0952	2.4	2100	1.6303	0.6639	1.9842	-1.3673	0.6444	3.3515	2.5293	1.6377	1.6287
0.1951	2.51	2200	1.1515	0.6784	-0.0674	-2.4660	0.6637	2.3985	2.4589	1.1463	1.1331
0.1119	2.63	2300	1.3845	0.6722	4.4149	1.2669	0.6548	3.1480	2.4797	1.3869	1.3759
0.1613	2.74	2400	1.1948	0.6536	-4.3162	-7.1133	0.6367	2.7971	2.4661	1.2014	1.1887
0.1408	2.86	2500	1.4167	0.6557	-3.1501	-6.3592	0.6415	3.2091	2.4591	1.4242	1.4137
0.2694	2.97	2600	1.2168	0.6536	0.5185	-2.2531	0.6395	2.7716	2.4397	1.2074	1.1949
0.1184	3.09	2700	1.6729	0.6412	0.5427	-3.2829	0.6315	3.8257	2.4188	1.6627	1.6551
0.1004	3.2	2800	1.8768	0.6742	3.9205	-0.6543	0.6629	4.5748	2.3906	1.8625	1.8572
0.1029	3.31	2900	1.7461	0.6619	0.1775	-4.2079	0.6496	4.3854	2.3534	1.7356	1.7294
0.0401	3.43	3000	1.9949	0.6825	3.6497	-1.3819	0.6698	5.0317	2.3327	1.9902	1.9868
0.04	3.54	3100	2.0206	0.6763	-0.5106	-5.0903	0.6597	4.5798	2.3202	2.0224	2.0194
0.1035	3.66	3200	2.1971	0.6660	2.3511	-2.5645	0.6536	4.9156	2.3137	2.2218	2.2209
0.0589	3.77	3300	2.1599	0.6412	2.0054	-2.7469	0.6262	4.7523	2.2936	2.1789	2.1777
0.084	3.89	3400	2.2096	0.6598	1.7952	-3.0061	0.6391	4.8013	2.2833	2.2386	2.2382
0.063	4.0	3500	2.2277	0.6660	4.2291	-0.8513	0.6484	5.0805	2.2693	2.2539	2.2537
0.065	4.11	3600	2.3431	0.6598	2.1719	-3.1923	0.6444	5.3642	2.2499	2.3575	2.3585
0.0453	4.23	3700	2.4069	0.6474	5.6839	0.2229	0.6335	5.4609	2.2344	2.4327	2.4347
0.0377	4.34	3800	2.4983	0.6557	2.7785	-2.9928	0.6355	5.7714	2.2258	2.5397	2.5429
0.0559	4.46	3900	2.4027	0.6536	2.8063	-2.8587	0.6375	5.6650	2.2135	2.4278	2.4299
0.0219	4.57	4000	2.4322	0.6598	3.9024	-1.8412	0.6435	5.7436	2.2081	2.4805	2.4832
0.09	4.69	4100	2.4041	0.6680	3.7769	-1.9890	0.6496	5.7659	2.2011	2.4248	2.4271
0.0897	4.8	4200	2.3727	0.6722	2.7679	-3.0182	0.6524	5.7861	2.1974	2.3815	2.3833
0.0474	4.91	4300	2.4124	0.6660	3.4303	-2.3575	0.6484	5.7878	2.1950	2.4421	2.4446

Framework versions

Transformers 4.37.2
Pytorch 2.4.0+cu121
Datasets 2.21.0
Tokenizers 0.15.2

Type: opt
License: other
Created: Sep 16, 2024
Updated: Sep 16, 2024