metadata

base_model: ondevicellm/tinyllama_moe_v2
tags:
  - alignment-handbook
  - generated_from_trainer
  - trl
  - sft
  - generated_from_trainer
datasets:
  - HuggingFaceH4/ultrachat_200k
model-index:
  - name: tinyllama_moe_sft_ultrachat_v2_ep3
    results: []

tinyllama_moe_sft_ultrachat_v2_ep3

This model is a fine-tuned version of ondevicellm/tinyllama_moe_v2 on the HuggingFaceH4/ultrachat_200k dataset. It achieves the following results on the evaluation set:

Loss: 1.1289

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: 2e-05
train_batch_size: 16
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 2
total_train_batch_size: 128
total_eval_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 120
num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.4892	0.09	100	1.4465
1.2729	0.18	200	1.2643
1.2286	0.26	300	1.2280
1.2007	0.35	400	1.2075
1.1688	0.44	500	1.1933
1.1872	0.53	600	1.1830
1.1732	0.61	700	1.1746
1.1596	0.7	800	1.1679
1.1546	0.79	900	1.1622
1.1366	0.88	1000	1.1572
1.1606	0.96	1100	1.1527
1.0967	1.05	1200	1.1505
1.099	1.14	1300	1.1480
1.1099	1.23	1400	1.1453
1.1015	1.31	1500	1.1432
1.104	1.4	1600	1.1408
1.0998	1.49	1700	1.1390
1.0829	1.58	1800	1.1369
1.1052	1.66	1900	1.1353
1.1082	1.75	2000	1.1336
1.0948	1.84	2100	1.1320
1.0682	1.93	2200	1.1308
1.0688	2.01	2300	1.1318
1.0754	2.1	2400	1.1317
1.0646	2.19	2500	1.1311
1.058	2.28	2600	1.1305
1.0553	2.37	2700	1.1301
1.0607	2.45	2800	1.1298
1.0669	2.54	2900	1.1294
1.0476	2.63	3000	1.1292
1.0688	2.72	3100	1.1291
1.0583	2.8	3200	1.1289
1.0545	2.89	3300	1.1289
1.0744	2.98	3400	1.1289

Framework versions

Transformers 4.36.2
Pytorch 2.0.1+cu117
Datasets 2.16.1
Tokenizers 0.15.0