sDarth/24b1064-Week08-Compression-40-Submission01

Week 08 Compression 40 — Submission 01

This submission recreates the best 40% experiment:

• source checkpoint: merged behavioral_concise_lora_r8 Qwen3.5-4B checkpoint
• quantizer: GPTQModel
• weights: W4A16
• group size: 64
• desc_act: false
• HMMT is never used for calibration

The earlier experiment with this recipe reached 55/100 on HMMT100 with 100% parse. The model weights themselves are not stored in git; build the packed checkpoint and upload it to the Week08 Hugging Face model repository.

1. Recreate/persist the behavioral BF16 parent

The expected parent checkpoint is:

/marimo/CS6013/artifacts/behavioral_concise_lora_r8/bf16_model

Use the repository's configs/distill/qwen35_4b_behavioral_concise_lora_r8.yaml and verified concise SFT traces to rebuild it if necessary. Freeze/save that parent checkpoint before quantization.

2. Compress

Install dependencies:

cd /marimo/CS6013/24b1064/Week08/Compression_40/Submission01
uv sync

Run GPTQ W4/G64 using a held-out math calibration JSONL (prompt + completion; no HMMT):

uv run python compress.py \
  --model_name Qwen/Qwen3.5-4B \
  --checkpoint_path /marimo/CS6013/artifacts/behavioral_concise_lora_r8/bf16_model \
  --calibration_path /marimo/CS6013/data/traces/behavioral_concise_lora_r8_rac_150k.jsonl \
  --output_path /marimo/CS6013/artifacts/week08_behavioral_gptq_w4_g64/compressed

The compressor writes compression_recipe.json and refuses success if the measured text-model fraction is not < 0.40 relative to the course 8.0585 GiB text baseline.

3. Decompress

uv run python decompress.py \
  --model_name Qwen/Qwen3.5-4B \
  --checkpoint_path /marimo/CS6013/artifacts/week08_behavioral_gptq_w4_g64/compressed \
  --output_path /marimo/CS6013/artifacts/week08_behavioral_gptq_w4_g64/decompressed

The output is a normal dense Hugging Face checkpoint for vLLM evaluation.

4. Submission model repo

The uploaded artifact is in the public Hugging Face repository:

sDarth/24b1064-Week08-Compression-40-Submission01

Upload the complete packed compressed/ directory only after the hard size check passes. Also preserve compression_recipe.json, the training config, calibration metadata, and the behavioral parent checkpoint/adapter in durable storage.

The Week08 evaluator audit measured size_frac=0.3705326200. The existing reconstructed-checkpoint compatibility smoke passed 3/3 problems with 100% parse rate. Fresh Python 3.12 submission dependency installation and imports also passed. These checks do not represent a new full HMMT evaluation.