svd-safety-llama3_8b_instruct_up_svdllm_finetuned_keep_0p60

llama3_8b_instruct_up_svdllm_finetuned_keep_0p60

meta-llama/Meta-Llama-3-8B-Instruct compressed with the complete SVD-LLM method using the authors' code (AIoT-MLSys-Lab/SVD-LLM, 7538cca98880), with one necessary change.

Grouped-query patch. Upstream's SVD_LlamaAttention assumes multi-head attention, so it cannot run a forward pass on Llama-3.1 (32 query heads, 8 KV heads) and its LoRA stage fails outright. component/svd_llama.py was patched to size k/v by num_key_value_heads and apply repeat_kv, both lifted from upstream's own component/svd_mistral.py (Mistral is grouped-query too), plus a causal-mask slice for transformers >=4.43. Patched file sha256 e65f644f9316. The truncation arithmetic is unchanged -- all 224 layer ranks are identical to those the unpatched clone produces, since whitening() derives every factor from the real weight shapes. Only the module's declared shapes differ.

• Parameters removed: 40% (keeping 60%)
• Realised parameter fraction: 0.5997971754807693
• Pipeline: data whitening -> SVD truncation -> LoRA on the U factors -> merge -> LoRA on the V factors -> merge -> fold to a dense checkpoint
• Whitening calibration: 256 WikiText-2 sequences of 2048 tokens, seed 3
• Parameter update: LoRA r=8, 2 epochs per factor, lr 0.0001, batch 64, on yahma/alpaca-cleaned

The factors are folded back to dense Llama shapes (W = U @ V), so this loads with stock transformers and needs no custom modelling code. It is rank-deficient, not smaller on disk.

Results

Over-refusal judging is reliable for this cell (XSTest-safe scored fraction 0.98; OR-Bench-Hard-1K scored fraction 0.97).

Full per-prompt outputs and raw metric files are in utility/ and safety/.

Evaluation

WikiText-2 perplexity; ARC-Easy/Challenge, HellaSwag, WinoGrande, OpenBookQA, PIQA and MathQA zero-shot; AdvBench and StrongREJECT generation judged by cais/HarmBench-Llama-2-13b-cls; over-refusal on XSTest-safe and OR-Bench-Hard-1K judged by allenai/wildguard. All generation uses the chat template, greedy.

Caveat

Compression at this ratio can degrade generation quality. Safety numbers from a model that has become degenerate are not evidence about alignment -- read the over-refusal and instruction-following columns alongside the ASR before drawing behavioural conclusions.