anjohn0077/NEXS-toxicity-lora

NEXS toxicity LoRA (vLLM-ready)

Rank-128 LoRA adapter for the toxicity domain, extracted with mergekit from K-intelligence/Llama-SafetyGuard-Content-Binary against the base model meta-llama/Llama-3.1-8B, then sanitized for vLLM serving.

Sanitization applied

The raw mergekit extraction included full-rank modules_to_save tensors (embed_tokens, lm_head, and RMSNorm layers) that vLLM's LoRA runtime does not support. This upload contains only the pure low-rank lora_A/lora_B weights (224 pairs: 32 layers x q/k/v/o/gate/up/down projections, bf16), with modules_to_save: null in adapter_config.json. No resize_token_embeddings() call is needed to load this adapter.

Serving with vLLM

python -m vllm.entrypoints.openai.api_server \
    --model meta-llama/Llama-3.1-8B \
    --enable-lora \
    --lora-modules toxicity=anjohn0077/NEXS-toxicity-lora \
    --port 8000 \
    --max-lora-rank 128 \
    --gpu-memory-utilization 0.85

Evaluation (sst2)

Variant	Accuracy
Base model	0.6732
This LoRA on base (via vLLM)	0.8991
Original full fine-tune	0.8899

Evaluated with lm-evaluation-harness against a local vLLM OpenAI-compatible endpoint:

lm_eval --model local-completions \
    --model_args model=toxicity,base_url=http://localhost:8000/v1/completions,tokenizer=meta-llama/Llama-3.1-8B,num_concurrent=10 \
    --tasks sst2 \
    --output_path results/vllm_toxicity