litfury/25M0831-Week01-Track1-Submission01

25M0831 — Week01 — Track 1 — Submission 01 (40% target)

Compressed checkpoint for CS6013 (Efficient AI). Not directly loadable via from_pretrained — this is a packed-int4 payload, not a standard HF checkpoint. Restore it first using the code in the companion GitHub repo.

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Technique: INT4 group-wise weight-only quantization (round-to-nearest, group size 128)
• Compressed size: 2.53 GiB (32.9% of the 7.68 GiB original)
• GitHub repo (compression/restore code): https://github.com/mangeshpoojan/CS6013/tree/main/25M0831/Week01/Track_1/Submission01

Restoring this checkpoint

```bash git clone https://github.com/mangeshpoojan/CS6013 cd CS6013/25M0831/Week01/Track_1/Submission01

python dequantize_to_bf16.py \ --model Qwen/Qwen3-4B-Instruct-2507 \ --checkpoint <path-to-downloaded-compressed.pt> \ --output ./restored_hf ```

`restored_hf/` is then a standard bf16 Hugging Face checkpoint, loadable via `AutoModelForCausalLM.from_pretrained("./restored_hf")`.