Reduce C4 load time with subset from Nota's S3

[bokyeong1015]

Purpose of This PR

Improve C4 data loading speed (used as calibration data for GPTQ).

• C4 load time on Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz:
  • Reduced from 1784s to 1s.

Changes in src/dataset.py

• Removed slow resolving of data en/c4-train.00000-of-01024.json.gz.
• Now, a subset of the C4-train first split is downloaded from Nota's S3 bucket:
  • 256 randomly selected long sequences (with length > 3072).
  • To see the text data, please visit this link.

Reproducibility Check

• CPT-Shortened-Vicuna-GPTQ (4-bit weight quantization)
  • The GPTQ-and-evaluation script remains unchanged.

Model Size	PPL-Wiki2	PPL-PTB	ACC-AVG	BoolQ	PIQA	HellaSwag	WinoGrande	ARC-e	ARC-c	OBQA
5.5B (Paper)	15.1	59.3	60.6	69.7	75.9	68.9	63.9	68.5	38.5	38.6
5.5B (PR #11)	15.2	59.9	60.9	69.7	75.5	69.1	65.0	69.0	39.0	39.0
5.5B (This PR)	15.2	59.4	60.8	70.4	75.2	69.1	64.3	68.4	39.2	39.2
3.7B (Paper)	16.6	61.5	57.1	63.8	74.5	62.7	61.0	65.8	34.2	37.8
3.7B (PR #11)	16.5	60.8	56.9	64.3	74.2	62.4	61.2	65.7	33.0	37.2
3.7B (This PR)	16.6	61.9	56.7	63.1	74.5	62.6	60.9	65.6	33.9	36.6
2.7B (Paper)	17.7	64.7	54.6	61.9	73.1	58.4	58.8	62.5	31.8	35.6
2.7B (PR #11)	17.7	64.5	54.5	62.1	73.3	58.7	57.1	61.4	30.3	38.2
2.7B (This PR)	17.7	65.2	54.4	61.5	73.0	58.1	58.0	62.4	30.6	36.8
1.5B (Paper)	21.4	80.0	48.5	48.9	70.1	48.8	54.1	55.7	26.8	35.0
1.5B (PR #11)	21.1	80.1	48.8	52.5	70.1	48.5	54.6	56.2	27.4	32.6
1.5B (This PR)	21.3	80.0	48.5	50.9	69.9	48.7	54.2	55.9	27.0	32.8