| --- |
| base_model: meta-llama/Llama-3.2-1B-Instruct |
| library_name: peft |
| tags: [lora, quantization, safety] |
| --- |
| |
| # Llama-3.2-1B-Instruct — configuration-matched GSM8K LoRA adapters |
|
|
| Five LoRA adapters, each trained on GSM8K under a **different quantization |
| configuration** of the same base model, for the Phase 1 transfer study. |
|
|
| | subfolder | base configuration it was trained under | |
| | --- | --- | |
| | `q16/` | bf16 (no quantization) | |
| | `q8/` | c1sim W8A16 | |
| | `q4/` | c2sim W4A16 | |
| | `q3/` | c5 W3A16 | |
| | `q2/` | c6 W2A16 | |
|
|
| r=16, α=32, dropout 0.05, bf16, on the 7 attention+MLP projections. Single seed |
| (0); see `PREREGISTRATION.md` §5 for why, and what it costs. |
|
|
| ## The result these exist to demonstrate |
|
|
| Transfer is **strongly asymmetric in direction**. An adapter trained at 3-bit transfers |
| *upward* nearly intact, but adapters trained at high precision, deployed *downward* at |
| 3-bit, score **below the no-adapter baseline** — fine-tuning on a clean configuration |
| actively hurts you at low precision: |
|
|
| | | deploy bf16 | deploy W4 | deploy W3 | |
| | --- | --- | --- | --- | |
| | no adapter | 0.313 | 0.315 | 0.049 | |
| | trained bf16 | 0.375 | 0.290 | **0.019** | |
| | trained W3 | 0.337 | 0.307 | **0.288** | |
|
|
| GSM8K exact match. Full matrices and the pre-registered asymmetry test: |
| [`Jeesup/safety-quant-phase1`](https://huggingface.co/datasets/Jeesup/safety-quant-phase1). |
|
|
| ## Use |
|
|
| ```python |
| from peft import PeftModel |
| model = PeftModel.from_pretrained(base, "Jeesup/Llama-3.2-1B-Instruct-safetyquant-lora", subfolder="q4") |
| ``` |
|
|
| Match the subfolder to the configuration you are actually deploying under. |
|
|
