PEFT Fine-Tuning of LLMs: LoRA, QLoRA, AdaLoRA

Problem: LLM not fitting your task?

You fine-tuned Llama 3.1 8B on general data, but it gives poor answers to questions about your legal documentation. Full fine-tuning requires 80 GB VRAM and several A100s—not everyone has those resources. Parameter-Efficient Fine-Tuning (PEFT) solves this: only 0.1–5% of parameters are updated, while the base model stays frozen. We use PEFT to adapt any LLM (from LLaMA to Mistral) to your task—classification, generation, RAG—without extra GPUs.

Our experience in LLM fine-tuning spans 5+ years, with hundreds of projects for clients in FinTech and LegalTech. We guarantee stable accuracy and optimized inference.

What problems does PEFT solve?

• Lack of GPU memory. Full fine-tuning of a 7B model requires ~56 GB VRAM with Adam. LoRA with rank 16 reduces it to ~18 GB, QLoRA (4-bit) to ~9 GB. One A100 instead of a cluster.
• Long training times. Full fine-tuning of Llama 3.1 8B takes 8 hours on 4×A100. LoRA r=16 takes 55 minutes on 1×A100—an 8–10x difference.
• Catastrophic forgetting. Full fine-tuning often degrades generalization. PEFT preserves base knowledge by only training the adapter.

Which PEFT method to choose for your task?

Selection depends on data volume, resources, and latency requirements. Here's a summary:

As noted in the Hugging Face PEFT documentation, PEFT can reduce resource requirements by 90%.

LoRA: the golden standard of PEFT

LoRA (Low-Rank Adaptation) adds low-rank matrices to attention layers (r=8–16). After training, they are merged with base weights—no inference latency increase. Example configuration:

from peft import LoraConfig, get_peft_model

config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    task_type="CAUSAL_LM",
)
model = get_peft_model(model, config)

On a financial news sentiment classification task (1200 examples, Llama 3.1 8B), LoRA r=16 achieved accuracy 0.91 vs. 0.74 for 5-shot without fine-tuning. Full fine-tuning gave 0.93 but took 8x longer.

AdaLoRA: adaptive rank for complex cases

AdaLoRA automatically distributes the parameter budget across layers, assigning higher rank to more important ones. Useful when layer criticality is unknown.

from peft import AdaLoraConfig, get_peft_model

config = AdaLoraConfig(
    init_r=12,
    target_r=8,
    beta1=0.85,
    beta2=0.85,
    deltaT=10,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
    task_type="CAUSAL_LM",
)
model = get_peft_model(model, config)

Prefix Tuning and IA³: when data is very scarce

Prefix Tuning adds learnable virtual tokens (20–100), IA³ uses scaling vectors. Both require <0.1% parameters and work with 50–200 examples. LoRA outperforms them 2–3x in accuracy when 500+ examples are available.

from peft import PrefixTuningConfig

config = PrefixTuningConfig(
    task_type="CAUSAL_LM",
    num_virtual_tokens=20,
    prefix_projection=True,
)
model = get_peft_model(model, config)

How we fine-tuned Llama 3.1 for financial news sentiment analysis

Task: Classify sentiment (Positive/Negative/Neutral) from 1200 texts. Base model: Llama 3.1 8B Instruct.

Verdict: LoRA r=16 is the best accuracy/resource trade-off. QLoRA gives comparable accuracy with half the VRAM.

Managing multiple adapters with PEFT

The PEFT library allows loading and switching adapters in a single base model:

from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
model = PeftModel.from_pretrained(base_model, "./adapter-legal", adapter_name="legal")
model.load_adapter("./adapter-finance", adapter_name="finance")
model.load_adapter("./adapter-medical", adapter_name="medical")

model.set_adapter("legal")
output_legal = model.generate(...)

This architectural pattern—one base instance, multiple specializations—reduces memory consumption by 5x.

What's included in the work

• Analysis of your task and selection of PEFT method.
• Dataset preparation and labeling (if needed).
• Training with metric monitoring (loss, accuracy, F1).
• Testing on a holdout set, comparison with baseline.
• Integration of the adapter into your pipeline (Hugging Face, SageMaker, Triton).
• Documentation and team training.
• Post-training support and fine-tuning adjustments as data evolves.

Timelines

• PEFT method selection and experiments: 3–7 days.
• Data preparation: 2–4 weeks.
• Training and method comparison: 1–2 weeks.
• Total: 3–6 weeks. We'll refine the timeline for your project during a consultation.

Contact us for a free analysis of your task—we'll select the optimal PEFT method and provide a detailed fine-tuning plan. Request a consultation to get an individualized plan.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.