distily/short_gpt2

Summary

Distilled with Distily library using teacher model gpt2 on dataset wikimedia/wikipedia.

Model Architecture:

• Architecture: GPT2LMHeadModel
• Total Parameters: 81,912,576
• Data Type (dtype): torch.bfloat16
• Model Size: 0.16 GB

Evaluation Metrics Comparison

step	epoch	enwikippl	frwikippl	loss	runtime	samples_per_second	steps_per_second	tinystoriesppl	zhwikippl
teacher eval		43.25	61.25					11.6875	19.125
0	0	2018634629120.0	122045790683136.0	21.0022	102.1494	97.896	12.237	9999220736.0	43705587204096.0
2500	0.0101	299008.0	6422528.0	5.8065	101.9861	98.053	12.257	45824.0	14483456.0
5000	0.0202	6880.0	96256.0	3.3113	102.9516	97.133	12.142	4160.0	493568.0
7500	0.0303	1216.0	8096.0	2.1560	103.0236	97.065	12.133	692.0	42752.0
10000	0.0404	608.0	3664.0	1.7825	102.3752	97.68	12.21	388.0	888.0
12500	0.0505	358.0	1632.0	1.4664	102.1871	97.86	12.232	272.0	308.0
15000	0.0606	288.0	1176.0	1.3488	102.6007	97.465	12.183	228.0	260.0
17500	0.0707	255.0	1040.0	1.2932	102.1542	97.891	12.236	199.0	215.0
20000	0.0808	216.0	892.0	1.1570	102.1073	97.936	12.242	173.0	149.0
22500	0.0909	178.0	740.0	1.0350	102.0765	97.966	12.246	146.0	141.0
25000	0.1010	155.0	524.0	0.9676	102.1019	97.941	12.243	122.5	139.0
27500	0.1111	142.0	560.0	0.9230	102.0256	98.015	12.252	114.0	130.0
30000	0.1212	137.0	470.0	0.8998	102.3365	97.717	12.215	108.5	138.0
32500	0.1313	134.0	476.0	0.8740	102.3911	97.665	12.208	104.0	140.0
35000	0.1414	129.0	496.0	0.8657	102.2153	97.833	12.229	102.5	141.0
37500	0.1515	127.0	464.0	0.8513	102.0489	97.992	12.249	97.0	117.0
40000	0.1616	108.0	446.0	0.7522	102.9331	97.15	12.144	93.0	104.0
42500	0.1717	99.5	374.0	0.6850	103.1088	96.985	12.123	82.0	116.0
45000	0.1818	90.5	346.0	0.6316	102.7903	97.285	12.161	73.5	113.0
47500	0.1919	82.5	320.0	0.5960	102.5988	97.467	12.183	71.0	101.0
50000	0.2020	78.5	306.0	0.5676	102.5936	97.472	12.184	72.5	106.0
52500	0.2121	79.5	290.0	0.5424	102.5863	97.479	12.185	64.5	92.0
55000	0.2222	76.0	270.0	0.5280	102.6307	97.437	12.18	65.0	87.0
57500	0.2323	76.5	272.0	0.5278	101.9639	98.074	12.259	64.5	102.0
60000	0.2424	77.5	268.0	0.5286	102.0921	97.951	12.244	62.75	99.5
62500	0.2525	75.5	264.0	0.5204	102.0679	97.974	12.247	63.25	83.0
65000	0.2626	76.0	260.0	0.5176	102.1795	97.867	12.233	61.5	90.5
67500	0.2727	74.5	256.0	0.5112	102.5764	97.488	12.186	62.25	93.5
70000	0.2828	73.5	258.0	0.5128	101.9569	98.081	12.26	62.0	79.0
72500	0.2929	75.0	250.0	0.5053	101.9382	98.099	12.262	64.0	96.0
75000	0.3030	72.5	238.0	0.5068	102.0407	98.0	12.25	61.5	88.5
77500	0.3131	73.5	256.0	0.5085	102.0542	97.987	12.248	64.5	86.5
80000	0.3232	70.5	238.0	0.4699	102.4042	97.652	12.207	54.75	98.5
82500	0.3333	68.0	242.0	0.4574	102.2684	97.782	12.223	55.5	160.0
85000	0.3434	64.5	218.0	0.4490	102.3277	97.725	12.216	52.0	77.5
87500	0.3535	66.5	203.0	0.4394	102.1134	97.93	12.241	51.25	67.5
90000	0.3636	63.75	212.0	0.4310	102.0438	97.997	12.25	51.25	88.5
92500	0.3737	65.5	209.0	0.4262	101.9984	98.041	12.255	49.75	103.5
95000	0.3838	65.0	204.0	0.4274	102.0781	97.964	12.246	46.25	83.0
97500	0.3939	64.5	201.0	0.4192	102.0692	97.973	12.247	50.5	94.5
100000	0.4040	64.5	203.0	0.4207	102.1283	97.916	12.24	49.0	88.0
102500	0.4141	63.0	209.0	0.4184	102.224	97.824	12.228	48.0	125.0
105000	0.4242	62.75	193.0	0.4166	102.1918	97.855	12.232	46.0	76.0
107500	0.4343	62.75	197.0	0.4128	102.1719	97.874	12.234	47.0	113.0
110000	0.4444	64.5	191.0	0.4118	103.0992	96.994	12.124	49.0	82.0
112500	0.4545	65.0	213.0	0.4128	102.7296	97.343	12.168	47.0	111.5
115000	0.4646	68.5	207.0	0.4301	102.178	97.868	12.234	49.0	108.0
117500	0.4747	65.0	217.0	0.4372	102.2302	97.818	12.227	50.25	124.0
120000	0.4848	65.5	210.0	0.4351	102.2952	97.756	12.22	51.0	139.0
122500	0.4949	66.0	272.0	0.4352	102.1941	97.853	12.232	50.5	226.0
125000	0.5051	67.0	240.0	0.4387	101.978	98.06	12.258	49.0	71.0
127500	0.5152	66.5	224.0	0.4396	101.9014	98.134	12.267	49.75	100.0
130000	0.5253	65.5	227.0	0.4354	102.1244	97.92	12.24	50.75	146.0
132500	0.5354	66.0	209.0	0.4286	102.0218	98.018	12.252	52.25	101.5
135000	0.5455	64.5	220.0	0.4361	101.9074	98.128	12.266	51.25	181.0
137500	0.5556	66.5	223.0	0.4288	102.0744	97.968	12.246	49.0	103.0
140000	0.5657	66.5	232.0	0.4287	102.1162	97.928	12.241	49.25	127.5
142500	0.5758	66.5	220.0	0.4299	101.9461	98.091	12.261	49.5	88.5
145000	0.5859	65.5	217.0	0.4238	101.9572	98.08	12.26	48.75	177.0
147500	0.5960	64.0	205.0	0.4109	101.9497	98.088	12.261	48.75	128.0
150000	0.6061	63.5	224.0	0.4051	102.0205	98.02	12.252	48.5	117.5
152500	0.6162	63.25	202.0	0.4000	101.9318	98.105	12.263	47.5	160.0
155000	0.6263	63.75	195.0	0.4052	102.0203	98.02	12.252	48.75	100.0
157500	0.6364	63.75	212.0	0.4014	101.8935	98.142	12.268	49.25	113.0
160000	0.6465	62.75	198.0	0.3988	101.9178	98.118	12.265	44.5	132.0
162500	0.6566	64.5	192.0	0.3918	102.0303	98.01	12.251	45.5	100.0
165000	0.6667	62.5	202.0	0.3958	102.3627	97.692	12.211	47.75	88.5
167500	0.6768	62.5	191.0	0.3883	102.1537	97.892	12.236	44.75	80.5
170000	0.6869	63.5	195.0	0.3880	102.0728	97.969	12.246	51.0	91.5
172500	0.6970	60.75	201.0	0.3863	101.9235	98.113	12.264	47.5	90.5
175000	0.7071	61.5	189.0	0.3806	101.9376	98.099	12.262	46.5	82.5
177500	0.7172	58.75	171.0	0.3512	101.9844	98.054	12.257	42.75	66.0
180000	0.7273	55.5	161.0	0.3218	101.881	98.154	12.269	39.25	54.0
182500	0.7374	54.25	149.0	0.3148	101.9839	98.055	12.257	38.75	47.75
185000	0.7475	53.5	160.0	0.3133	101.9875	98.051	12.256	38.75	45.0
187500	0.7576	54.75	160.0	0.3114	101.9762	98.062	12.258	38.0	43.75
190000	0.7677	53.75	147.0	0.3075	101.9972	98.042	12.255	38.0	38.25
192500	0.7778	54.0	157.0	0.3057	101.9431	98.094	12.262	38.0	48.0
195000	0.7879	53.25	149.0	0.3058	101.9778	98.061	12.258	37.0	41.0
197500	0.7980	54.0	152.0	0.3032	102.0059	98.034	12.254	37.25	40.0
200000	0.8081	53.75	151.0	0.3033	102.0615	97.98	12.248	37.25	47.25
202500	0.8182	53.0	146.0	0.2957	102.0116	98.028	12.254	36.75	39.0
205000	0.8283	52.5	139.0	0.2903	102.1449	97.9	12.238	36.5	35.75
207500	0.8384	52.0	142.0	0.2894	102.0126	98.027	12.253	36.25	38.25
210000	0.8485	52.25	142.0	0.2883	102.0938	97.949	12.244	36.0	37.25
212500	0.8586	52.5	141.0	0.2874	101.9515	98.086	12.261	36.0	37.0
215000	0.8687	52.25	140.0	0.2873	101.9427	98.094	12.262	36.0	36.0
217500	0.8788	51.75	141.0	0.2863	102.0114	98.028	12.254	36.0	35.5
220000	0.8889	52.0	141.0	0.2854	102.0424	97.999	12.25	36.0	35.75
222500	0.8990	52.5	143.0	0.2853	102.0368	98.004	12.25	36.0	35.25
225000	0.9091	52.0	142.0	0.2849	102.115	97.929	12.241	35.75	35.0
227500	0.9192	52.0	141.0	0.2851	102.0455	97.996	12.249	36.0	35.25
230000	0.9293	52.0	141.0	0.2846	102.0273	98.013	12.252	35.75	35.25
232500	0.9394	52.0	141.0	0.2843	101.961	98.077	12.26	35.75	35.0
235000	0.9495	52.0	141.0	0.2844	102.0188	98.021	12.253	35.75	35.25
237500	0.9596	52.0	141.0	0.2845	102.0714	97.971	12.246	35.75	35.25
240000	0.9697	52.0	141.0	0.2844	102.0371	98.004	12.25	35.75	35.25
242500	0.9798	52.0	141.0	0.2844	102.0363	98.004	12.251	35.75	35.25
245000	0.9899	52.0	141.0	0.2844	102.0254	98.015	12.252	35.75	35.25
247500	1.0	52.0	141.0	0.2846	102.5728	97.492	12.186	35.75	35.25

Resource Usage Comparison

• VRAM Use: 7.2012 GB

`# Distillation (Teacher -> Student) Architecture Difference:

• Architecture: GPT2LMHeadModel -> GPT2LMHeadModel
• Total Parameters: 124,439,808 -> 81,912,576
• Data Type (dtype): 124439808 -> torch.bfloat16
• Model Size: 0.24 GB -> 0.16 GB

Module Diff Details

--- teacher model modules
+++ student model modules
@@ -4,7 +4,7 @@
     (wpe): Embedding(1024, 768)
     (drop): Dropout(p=0.1, inplace=False)
     (h): ModuleList(
-      (0-11): 12 x GPT2Block(
+      (0-5): 6 x GPT2Block(
         (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
         (attn): GPT2FlashAttention2(
           (c_attn): Conv1D()

Train Dataset

Trained on 521,350,663 tokens from the wikimedia/wikipedia dataset.

• Num Samples: 990,000
• Subset: 20231101.en
• Split: train

Training Objective

DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl))

Hyperparameters

The following hyperparameters were used during training:

Expand

• learning_rate: 0.0001
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.5
• num_epochs: 1.0
• distillation_objective: DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl))
• train_embeddings: True
• lr_scheduler: <torch.optim.lr_scheduler.LambdaLR object at 0x7fd9b01df220>
• student_model_name_or_path: None
• student_config_name_or_path: distilbert/distilgpt2
• student_model_config: None
• reinitialize_weights: None
• copy_teacher_modules: [('lm_head', False)]
• student_model_as_bitnet: False
• student_model_compile: False
• dropout: None
• teacher_model_name_or_path: gpt2
• teacher_load_in_8bit: False
• teacher_load_in_4bit: False
• teacher_model_compile: False
• dataset_uri: wikimedia/wikipedia
• dataset_subset: 20231101.en
• dataset_split: train
• dataset_column_name: text
• dataset_sample_size: 1000000
• dataset_test_size: 0.01
• gradient_accumulation_steps: 1
• weight_decay: 0.0
• max_grad_norm: 1.0
• warmup_ratio: 0.5
• warmup_steps: 0
• gradient_checkpointing: True

Framework Versions

• Distily 0.2.0
• Transformers 4.44.0
• Pytorch 2.3.0
• Datasets 2.21.0