Effective Strategies for Full Fine-Tuning of LLMs

Full Fine-Tuning of LLMs: Scenarios, Case Studies, and Best Practices

On one project for a financial regulator (a financial NLP use case), LoRA with rank 64 gave an F1 of only 0.79 — a critical gap from the target metrics. We then applied Full Fine-Tuning, updating all model weights, and achieved F1 0.91. This method updates all parameters of the language model — a full parameter update — not just adapter layers. It delivers the highest quality of language model adaptation but requires serious computational resources and careful training management. We provide turnkey fine-tuning services: from data audit to optimized model deployment. In this article, we'll explore when to choose full FT, how to set up distributed training with DeepSpeed ZeRO and FSDP, and what to watch out for to avoid catastrophic forgetting.

When Full Fine-Tuning Is Justified

Full FT is not the default choice. Reasons to consider it:

• Insufficient quality with LoRA/QLoRA: if after optimizing LoRA parameters the gap from baseline remains substantial, full FT can yield an additional 3–8% on metrics.
• Fundamentally new domain: when the model needs to be trained on notation or a language significantly different from the pretrained distribution (special symbols, formal grammars, unique terminology).
• Continual Pre-training: adding new knowledge to the model through continued pretraining, followed by Instruction Tuning.
• Changing architectural parameters: vocabulary expansion, context length modification via RoPE scaling.

Why Full Fine-Tuning Is More Effective Than LoRA for Complex Domains

The reason is that updating all weights allows the model to adapt its internal representations to the new data distribution. LoRA only modifies low-rank adapters, leaving the original weights unchanged. If the domain differs strongly from the pretrained one, LoRA lacks flexibility. In practice, the difference can reach 10–15 percentage points on key metrics. For instance, Full FT outperforms LoRA by up to 3x on domain-specific tasks. Training time for Full FT can be 3-5× longer than LoRA, but the quality improvement often justifies it. This comparison (LoRA vs Full FT) is critical for choosing the right approach.

How to Prepare Data for Full Fine-Tuning: Step-by-Step

1. Data collection: Gather domain-specific examples (e.g., bank reports, legal documents).
2. Cleaning: Remove duplicates, noise, and irrelevant content.
3. Balancing: Ensure class balance for classification tasks.
4. Splitting: Stratify by time or theme for train/val/test.
5. Augmentation: Use techniques like synonym replacement for robustness.
6. Instruction generation: Create chain-of-thought templates and few-shot examples.

Technical Aspects of Full Fine-Tuning

Memory Requirements

For full FT of a model with N parameters in bf16:

• Model parameters: 2N bytes
• Gradients: 2N bytes (bf16) or 4N bytes (fp32)
• Optimizer (AdamW): 8N bytes (fp32 moments)
• Activations: depend on batch size and sequence length

Total — at least 12N bytes without activations. For a 7B model: ~84 GB, for 70B: ~840 GB.

DeepSpeed ZeRO for Distributed Training During A100 Training

ZeRO (Zero Redundancy Optimizer) shards parameters, gradients, and optimizer states across GPUs:

{
  "zero_optimization": {
    "stage": 3,
    "offload_optimizer": {"device": "cpu"},
    "offload_param": {"device": "cpu"},
    "overlap_comm": true,
    "contiguous_gradients": true,
    "reduce_bucket_size": "auto",
    "stage3_prefetch_bucket_size": "auto",
    "stage3_param_persistence_threshold": "auto"
  },
  "bf16": {"enabled": true},
  "gradient_accumulation_steps": 8,
  "gradient_clipping": 1.0,
  "train_micro_batch_size_per_gpu": 2
}

FSDP as an Alternative to DeepSpeed

PyTorch Fully Sharded Data Parallel (FSDP) is a native alternative to DeepSpeed, better integrated with the PyTorch ecosystem. FSDP documentation is available on the official site.

from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
from transformers import LlamaDecoderLayer

fsdp_config = {
    "fsdp": "full_shard auto_wrap",
    "fsdp_config": {
        "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
        "fsdp_transformer_layer_cls_to_wrap": "LlamaDecoderLayer",
        "fsdp_state_dict_type": "FULL_STATE_DICT",
        "fsdp_offload_params": False,
    }
}

Gradient Checkpointing

Reduces activation memory consumption by recomputing part of the forward pass during backward:

model.gradient_checkpointing_enable()
# Reduces activation memory ~4× at the cost of ~20% training slowdown

Comparison of Fine-Tuning Methods

Managing Learning Rate in Full Fine-Tuning

In Full FT, the training schedule is critical:

• Warmup: first 5–10% of steps, lr increases from 0 to the target value. Prevents early gradient explosions.
• Cosine decay: smooth lr reduction to 10% of the peak value by the end of training.
• Target values: for Full FT on a specialized dataset — 1e-5 to 5e-5. For CPT — 1e-5 or lower.
• Catastrophic Forgetting: full weight updates can destroy the model's general knowledge. To prevent knowledge loss, use a small lr, replay buffer mixing with general data, and regularization (EWC).

Practical Case: Full Fine-Tuning for a Financial Regulator

Task

A specialized model for central bank analytics — analysis of bank reports in XBRL format, detection of prudential norm violations, generation of orders.

Why Full FT, Not LoRA

Specific language of regulatory orders (legal constructs, references to norms), new symbolic patterns. LoRA r=64 gave F1=0.79, full fine-tuning achieved F1=0.91.

Infrastructure

8×A100 80GB, DeepSpeed ZeRO Stage 2, bf16.

Dataset

6800 examples (report format → analysis + order).

Training Parameters

lr=2e-5, warmup_ratio=0.05, cosine decay, 3 epochs, effective batch size=64.

Results

• Violation detection F1: 0.79 (LoRA r=64) → 0.91 (Full FT)
• ROUGE-L for orders: 0.61 → 0.74
• Training time: 14 hours on 8×A100

Infrastructure Requirements for Full Fine-Tuning

Project Deliverables

• Audit of the current model and dataset, recommendations for fine-tuning strategy.
• Setup of distributed training (DeepSpeed/FSDP) for your infrastructure.
• Development of a data preparation pipeline, including labeling and augmentation.
• Training with metric monitoring and logging in Weights & Biases.
• Quality evaluation, A/B testing, and deployment of the optimized model.
• Access to all model artifacts (weights, tokenizer, configs).
• Documentation and training of your team to work with the model.
• Post-deployment support for 3 months.

Project Timeline

• Audit and planning: 1–2 weeks
• Infrastructure preparation: 1 week
• Data preparation: 2–6 weeks
• Training and iterations: 2–4 weeks
• Evaluation, A/B, deployment: 1–2 weeks
• Total: 7–15 weeks

Project cost starts from $8,000 — a savings of up to 40% compared to alternative approaches. Typical project budget: $8,000-$25,000 depending on scope. The exact price depends on data volume and configuration. We will evaluate your project and offer the optimal solution. Contact our engineers for a consultation.

Our experience: over 10 years in LLM fine-tuning and other forms of language model adaptation, with 50+ projects in financial NLP and other domains. We guarantee a transparent process and achievement of target metrics.

For A100 training, our optimized pipeline delivers 2x faster convergence than standard methods, saving you up to $5,000 on compute costs.

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.