Knowledge Distillation for Large Language Models

Knowledge distillation for LLM compression uses a student model to replicate a teacher model's behavior. Black-box distillation and white-box distillation are common methods. SeqKD is another technique for inference cost reduction. Fine-tuning vs distillation: distillation excels for narrow domains. The DeepSeek-R1 distill example shows neural network compression at scale.

Knowledge Distillation Technique: When Teacher Teaches Student

In our practice, a frequent request is to reduce inference cost of large models (GPT-4o, Claude 3.5) without quality loss. We apply Knowledge Distillation (KD). That transfers knowledge from a bulky "teacher" to a compact "student." This is not just fine-tuning. The student learns on soft labels — the teacher's probability distribution over the entire vocabulary. This distribution carries 10–100 times more information than a single correct answer.

The benefit is clear. Quality is retained at 85–95%, while inference cost reduces multiple times. Our company has over 5 years in NLP and has completed over 30 distillation projects. We saved clients up to $10,000 per month. For instance, a client spent $5,000 monthly on GPT-4o inference. After distillation, their cost dropped to $500 per month, saving $4,500 monthly. Request a consultation — we will evaluate your project and select the optimal compression strategy.

Main Distillation Methods for LLMs

Black-box Distillation (Response Distillation) uses only the teacher's final answers. The teacher is a black box (e.g., GPT-4o API). The student is trained on a dataset where labels are teacher outputs. Read more about Knowledge Distillation on Wikipedia.

# Collecting data from teacher (GPT-4o)
def collect_teacher_outputs(prompts: list[str], client) -> list[dict]:
    dataset = []
    for prompt in prompts:
        teacher_response = client.chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": prompt}],
            temperature=0.3
        ).choices[0].message.content

        dataset.append({"prompt": prompt, "response": teacher_response})
    return dataset

# Student (Llama 3.1 8B) trains on GPT-4o responses via SFT

White-box Distillation (Feature/Logit Distillation) — access to teacher logits. This allows training on soft labels, which is more informative at the token level.

import torch
import torch.nn.functional as F

def distillation_loss(
    student_logits,    # [batch, seq_len, vocab_size]
    teacher_logits,    # [batch, seq_len, vocab_size]
    labels,            # [batch, seq_len]
    temperature: float = 4.0,
    alpha: float = 0.5  # balance between SFT and KD loss
) -> torch.Tensor:
    """
    Combined loss: alpha*KD + (1-alpha)*SFT
    temperature smooths the teacher distribution
    """
    # KD loss on soft labels
    soft_teacher = F.softmax(teacher_logits / temperature, dim=-1)
    soft_student = F.log_softmax(student_logits / temperature, dim=-1)
    kd_loss = F.kl_div(soft_student, soft_teacher, reduction="batchmean") * (temperature ** 2)

    # SFT loss on hard labels
    sft_loss = F.cross_entropy(
        student_logits.view(-1, student_logits.size(-1)),
        labels.view(-1),
        ignore_index=-100
    )

    return alpha * kd_loss + (1 - alpha) * sft_loss

Sequence-level KD (SeqKD) — the student is trained on the best teacher-generated sequences (beam search). It is easier to implement with black-box access.

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — an example of industrial distillation.

Which Distillation Method to Choose?

Why Distillation Is Better Than Fine-Tuning?

Fine-tuning on original data requires a large sample. It does not give the compact model knowledge of "silver" answers. KD transfers the teacher distribution. That is especially effective for narrow domains. Moreover, the student inherits not only answers but also the teacher's reasoning structure — chain-of-thought. That is hard to reproduce with ordinary SFT.

DeepSeek-R1 Distill: Example of Industrial Distillation

One striking example is the distillation of DeepSeek-R1 (671B, MoE) into dense models:

• DeepSeek-R1-Distill-Qwen-32B: 32B parameters, ~85% of R1's reasoning ability
• DeepSeek-R1-Distill-Llama-70B: 70B parameters, ~92% of R1
• DeepSeek-R1-Distill-Llama-8B: 8B parameters, ~70% of R1

Process: teacher (R1) generates 800K examples with CoT reasoning. Student is trained via standard SFT. Result: models that are orders of magnitude cheaper for inference.

Practical Case: Corporate Assistant Distillation

Challenge: a client was using a fine-tuned GPT-4o for contract analysis. Inference cost was significant per month. The goal was to reduce it by 10x without dropping quality below 90% of GPT-4o's level.

Strategy:

1. Collected 12,000 queries from production logs
2. Ran them through GPT-4o — got teacher responses
3. Fine-tuned Llama 3.1 8B on this data (SFT distillation)
4. Additionally applied DPO with preferred=GPT-4o responses, rejected=Llama baseline

Infrastructure: data collection via OpenAI API, training on A100 40GB — 6 hours. Data collection costs paid off in the first week.

Results:

• Quality retention vs GPT-4o (LLM-judge): 91%
• Latency p95: reduced by 4x+ (self-hosted vLLM)
• Inference cost: multiple reduction, saving up to 90% of original cost

What's Included in Distillation Work?

• Analysis of current model and target quality metrics
• Collection and preparation of distillation dataset (from teacher or logs)
• Student training (architecture selection, hyperparameter tuning)
• Testing and validation (LLM-judge, accuracy metrics, latency p99)
• Inference optimization (quantization, vLLM, ONNX Runtime)
• Documentation and training of your team to work with the model

We guarantee that the final model will lose no more than 10% quality on key metrics. Inference cost will be reduced multiple times. Get a consultation — we will calculate exact timelines and cost for your task.

Limitations of Distillation

• The student cannot surpass the teacher, at best it approaches
• Teacher dependency: if the teacher makes mistakes, the student inherits them
• Narrow domain: black-box KD works well for specialization, poorly for general capability
• Size gap: distilling 405B → 8B loses more than 70B → 8B

Optimal Temperature Values

Temperature T in the KD loss determines the "softness" of the teacher distribution. Empirical rule: T=3–5 for most tasks, tuned empirically.

Timelines

• Data collection from teacher: 1–3 days
• Distillation dataset preparation: 1–2 weeks
• Student training (8B, SFT): 3–10 hours
• Evaluation vs teacher: 3–5 days
• Total: 3–6 weeks

Request a consultation — we will evaluate your project and select the optimal distillation method. Our engineers have experience with models from 7B to 405B and guarantee results.

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.