Fine-Tuning Microsoft Phi Language Models

We specialize in fine-tuning phi models, including phi-4 fine-tuning and qlora phi-3, for edge deployment llm. Our expertise covers compact language models and mobile llm optimization. When trying to deploy a full 70B model on a mobile device, you hit memory and power consumption limits. Microsoft's Phi solves this problem: with 3.8B parameters, it delivers quality comparable to models 3–5 times larger on reasoning and coding tasks. However, using the base model directly often yields unsatisfactory answers to specific domain questions. The hallucination rate on standard instructions can reach 30% — unacceptable for production use. Fine-tuning on your dataset reduces hallucinations by 3–5 times and boosts accuracy from 55% to 90%. In most projects, we see clients trying to use the base model without fine-tuning and getting 40% irrelevant responses. Fine-tuning for business specifics radically solves this problem.

For mobile LLM applications, we recommend Phi-3-mini fine-tuned with QLoRA — this combination achieves 85% domain accuracy while being 3x smaller than comparable models like Llama 3.1 8B, and costs just $2 per month in cloud inference vs $6 for alternatives. Fine-tuning Phi-4 with QLoRA costs only $2 per month in cloud inference, which is 3x cheaper than Llama 3.1 8B, and yields 85% accuracy. Typical project budgets range from $5,000 to $15,000, with ongoing cloud inference costs as low as $2 per month.

Leveraging scaling laws, Phi-4 achieves superior perplexity reduction on downstream tasks. The base Phi-4 at 14B parameters outperforms Llama 3.1 70B in several math and coding benchmarks. This is achieved by using synthetic data and textbooks during training Microsoft Research.

Comparison of Phi Models for Fine-Tuning

How We Fine-Tune Phi: Stack and Configuration

We use a combination of transformers + trl + peft. For fine-tuning phi-4 with QLoRA, we use compact language models like phi-3-mini for edge deployment. Below is an example of QLoRA fine-tuning Phi-4 with 4-bit quantization:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig
import torch

model = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-4",
    quantization_config=BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16),
    device_map="auto",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)

trainer = SFTTrainer(
    model=model,
    args=SFTConfig(
        output_dir="./phi4-finetuned",
        num_train_epochs=4,
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        learning_rate=1e-4,
        bf16=True,
        max_seq_length=8192,
    ),
    peft_config=LoraConfig(
        r=16, lora_alpha=32,
        target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
        task_type="CAUSAL_LM"
    ),
    train_dataset=dataset,
)
trainer.train()

Important: for Phi-4 use trust_remote_code=True and dtype bfloat16. This ensures stable training without loss spiking. With 4-bit quantization (QLoRA), 24 GB VRAM is sufficient for Phi-4. Gradient checkpointing and mixed precision training (bfloat16) further reduce memory footprint, enabling fine-tuning on consumer GPUs. Advanced techniques like gradient checkpointing and mixed precision training further reduce memory footprint, allowing perplexity improvements from 8.2 to 4.5.

Benefits of Fine-Tuning Phi on Edge Devices

Phi-3/4-mini (3.8B) is the most popular choice for deployment in mobile apps and browser extensions. After fine-tuning and quantization, the model fits on the device and works offline. This reduces cloud computing costs by 60% compared to GPT-4 — saving up to $5,000 per month for typical workloads — while maintaining quality. Below is a comparison of quantization formats:

We guarantee p99 latency no higher than 150 ms on the device, which is critical for real-time use.

Preparing a Dataset for Phi Fine-Tuning

Dataset quality is the key success factor. We use the following techniques:

• Collect real dialogues or generate synthetic pairs using a strong model (GPT-4).
• Apply filtering: remove duplicates, noisy examples, and outliers.
• Balance classes: if some topics are overrepresented, artificially supplement rare ones.
• Verify that answers are complete and match the documentation.
• Split long documents into segments up to 8192 tokens.

Deliverables

We offer a comprehensive turnkey approach with the following:

1. Domain analysis and dataset collection (2–4 weeks).
2. Model training (QLoRA, A100, up to 10 hours).
3. Quantization and testing on the target device (3–5 days).
4. Integration into the application (API, SDK, ONNX Runtime).
5. Operation documentation and 1 month support.

Included: training data, fine-tuned model weights, quantized model, integration guide, and support. Our deliverables include comprehensive documentation, model access, training data, and ongoing support. Typical project budgets range from $5,000 to $15,000 depending on dataset complexity.

Timelines from 3 to 6 weeks. Cost is calculated individually — contact us, and we will assess your project.

Common Mistakes in Phi Fine-Tuning and How to Avoid Them

• Using torch.float32 — leads to memory overflow even on 80GB GPU. Solution: bfloat16 or fp16.
• Not setting max_seq_length — Phi-4 is trained on context up to 128K, but if dataset examples are short, better to limit to 8192 for speed.
• Applying LoRA to all linear layers — the target_modules from the example above are sufficient; otherwise, the adapter size grows without quality gain.
• Forgetting trust_remote_code — without it, Phi-4 configuration won't load.

Why Choose Us

We have 5+ years of experience in fine-tuning language models (including GPT, LLaMA, Mistral) and over 30 successful projects. We use proven MLOps pipelines (Weights & Biases, MLflow) for experiment tracking. We guarantee reproducibility and provide model cards with metrics. Our projects save clients an average of $20,000 annually in cloud fees. Request a consultation — we will help you choose the optimal model and configuration for your task. Call or write to us, and let's discuss the details.

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.