Dynamic Multi-Adapter Routing: LoRA Weights & Meta-Router Artifacts

This repository hosts the domain adapters, the TIES hybrid adapter, and routers

All adapters are trained on top of the meta-llama/Llama-3.2-3B base model.

• GitHub Repository: ozgur-coban/confidence-driven-lora
• Evaluation Datasets & Logs: ozgur-coban/multi-adapter-datasets

---

Repository Contents

Path / Filename	Type	Description
math_adapter_3b/	PEFT LoRA Adapter	Trained on 7,500 samples from openai/gsm8k for mathematical reasoning.
code_adapter_3b/	PEFT LoRA Adapter	Trained on 15,000 length-filtered Python samples from m-a-p/CodeFeedback-Filtered-Instruction.
hybrid_ties_code_math_3b/	PEFT LoRA Adapter	Offline TIES-Merged adapter combining math_adapter_3b and code_adapter_3b (density=0.8, weights 0.5 / 0.5).
syntax_router.joblib	Joblib Artifact	Pre-trained TF-IDF vectorizer (1–3 n-grams, 10k features) + Logistic Regression classifier.
semantic_centroids.npy	NumPy Artifact	Pre-calculated 768D domain centroids using sentence-transformers/all-mpnet-base-v2.
meta_router.joblib	Joblib Artifact	6-dimensional Random Forest Meta-Classifier (n_estimators=100, max_depth=4).

---

Training Hyperparameters

The single-domain adapters (math_adapter_3b and code_adapter_3b) were trained on a Google Colab NVIDIA T4 GPU using 4-bit NormalFloat (NF4) QLoRA via bitsandbytes with -100 loss masking applied to prompt tokens:

• Base Model: meta-llama/Llama-3.2-3B
• Rank ($r$): 32
• LoRA Alpha ($\alpha$): 64
• LoRA Dropout: 0.0
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Optimizer: adamw_torch
• Learning Rate: $3 \times 10^{-4}$ (Cosine scheduler, 50 warmup steps)
• Batch Size: 8 (Gradient Accumulation: 2, Effective Batch Size: 16)
• Max Sequence Length: 512
• Epochs: 1
• Precision: FP16 (torch.float16)

---

Usage: Loading Multi-Adapters with PEFT

You can mount all three adapters simultaneously into VRAM and dynamically switch between them in ~2.44 ms using set_adapter():

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from huggingface_hub import snapshot_download

# Download repository artifacts
repo_dir = snapshot_download(repo_id="ozgur-coban/multi-adapter")

base_model_id = "meta-llama/Llama-3.2-3B"
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load Math adapter first, then mount Code and Hybrid adapters
model = PeftModel.from_pretrained(
    base_model,
    f"{repo_dir}/math_adapter_3b",
    adapter_name="math"
)
model.load_adapter(f"{repo_dir}/code_adapter_3b", adapter_name="code")
model.load_adapter(f"{repo_dir}/hybrid_ties_code_math_3b", adapter_name="hybrid")

# Switch active adapter dynamically per query
model.set_adapter("math")
inputs = tokenizer(
    "Question: Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?\nAnswer:\n",
    return_tensors="pt"
).to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# Or disable adapters to use the raw base model for general queries
with model.disable_adapter():
    outputs = model.generate(**inputs, max_new_tokens=256)