Downloads · 30 days
0
Panzj/qwen0.5b-gsm8k-pcm-lora
qwen0.5b-gsm8k-pcm-lora is a machine learning model from Panzj. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Four GSM8K fine-tuned LoRA models for analog (PCM) noise/drift robustness research.
Downloads · 30 days
0
Access
Public
Updated Jul 24, 2026
Repo size
14.8 GB
Likes
1
Public
Click a slice to open those files.
.pt14.8 GB · 100%
From the Hugging Face model README
Four GSM8K fine-tuned LoRA models for analog (PCM) noise/drift robustness research.
| Model | W_res decomposition | Analog training | Size |
|---|---|---|---|
cheb/ | Chebyshev LRA (min-max) | ✅ 12ep | 3.3G |
pissa/ | SVD (Frobenius optimal) | ✅ 12ep | 3.3G |
lora_analog/ | None (pretrained) | ✅ 12ep | 3.3G |
lora/ | None (pretrained) | ❌ digital | 2.0G |
Shared: cheb_init.pt (1.4G) — Chebyshev LRA W_res for cheb model. eq_init_v3.pt (832M) — RMSNorm-folded gate/up variant.
conda create -n pissa-dora python=3.10
conda activate pissa-dora
pip install torch transformers peft datasets accelerate
pip install aihwkit # IBM analog hardware simulator
pip install huggingface_hub # for model download
python eval_noise.py --model cheb --noise 1.0
Runs all 1319 GSM8K test samples with batch=256. Options:
--model: cheb | pissa | lora_analog | lora (default: cheb)--noise: 0 (digital) | 1.0 | 2.0 (default: 1.0)--runs N: repeat N times for statistics (default: 1)python eval_drift.py --model cheb --time 1year --runs 3
Simulates conductance drift: g(t) = g₀×(t/20s)^(-ν), ν~0.06. Built-in GlobalDriftCompensation.
--time: t0 | 1min | 1hr | 1day | 1week | 1month | 1year | 10year--runs N: independent programming + drift repeats (default: 3)| Model | digital | n=1.0 | n=2.0 | 10yr drift |
|---|---|---|---|---|
| Cheb | 32.5% | 28.1% | 20.8% | 16.5% |
| PiSSA | 31.3% | 24.7% | 20.2% | 15.3% |
| LoRA+analog | 21.6% | 20.2% | 14.3% | 10.4% |
| LoRA (digital) | 27.1% | 1.1% | 1.0% | 1.0% |
If used in research, please cite the accompanying paper.