7B–70B+
Model Range
Tiered
Strategy (3 Tiers)
Provider-agnostic
Cloud GPU
Days
Dataset to Deploy
The Challenge
The client had AI use cases requiring custom model behavior - domain-specific knowledge, specialized response formats, particular reasoning patterns - that base models couldn't deliver. They needed a path from "we have a dataset" to "we have a production model" without building internal GPU infrastructure expertise or overspending on compute.
The key tension: the client's use cases ranged from rapid prototyping of small models to production fine-tuning of 70B parameter models. A single approach wouldn't work. A framework that matched method, hardware, and cost to each use case was required.
Strategic Framework Delivered
Tier 1: Rapid Prototyping (7B–13B Models)
| Aspect | Detail |
|---|---|
| GPU | Single RTX 4090 (24GB VRAM) |
| Method | QLoRA - 4-bit quantization enabling 7B models comfortably, 13B models workably |
| Framework | Unsloth - 2–5× faster than standard fine-tuning, significantly reduced VRAM usage |
| Cost | $2–8 per run (10k samples, 2–4 hours training) |
| Outcome | Working fine-tuned model prototypes in under one business day |
Tier 2: Production Fine-Tuning (13B–40B Models)
| Aspect | Detail |
|---|---|
| GPU | 2–4× A100 80GB or RTX 6000 Ada (96–192GB total VRAM) |
| Method | LoRA / QLoRA matched to model size and VRAM availability |
| Framework | Axolotl + DeepSpeed - production workloads, native FSDP support |
| Cost | $15–80 per run |
| Outcome | Production-quality fine-tuned models with full training metric history |
Tier 3: Large-Scale Training (70B Models)
| Aspect | Detail |
|---|---|
| GPU | 4–8× A100 80GB or H100 80GB clusters |
| Method | QLoRA or LoRA - full fine-tuning is not cost-effective at this scale |
| Framework | Axolotl + DeepSpeed ZeRO-3 with CPU/NVMe offloading |
| Cost | $50–200 per run |
| Outcome | 70B+ parameter models fine-tuned for production deployment at manageable cost |
Data Preparation Work
Jashom applied structured data preparation protocols to the client's training datasets before any GPU compute was spent:
- Consistency audit: enforced formatting standards across the full dataset
- Domain-expert validation pass: flagged low-quality or ambiguous examples for removal
- Deduplication: eliminated near-duplicate samples that inflate dataset size without adding learning signal
- Coverage analysis: identified gaps in use-case coverage and recommended targeted data collection
The principle applied throughout: model performance is determined more by data quality than volume. A 2,000-example high-quality dataset outperforms a 20,000-example noisy one.
Cloud Provider Strategy
Jashom's provider-agnostic approach matches cloud GPU availability to training needs at any given time. The client's workloads were executed across multiple providers depending on GPU availability, pricing, and reserved vs. spot instance tradeoffs:
- AWS (A100/H100 instances) - production runs with SLA requirements
- Lambda Labs and CoreWeave - cost-optimized long-form training runs
- RunPod - rapid prototyping and experimentation at low hourly cost
All runs configured with frequent checkpointing to S3/GCS and immediate instance termination after training - minimizing idle compute cost.
Technologies Used
Outcome – Strategic Value
The tiered framework delivered predictable training costs, faster iteration from dataset to deployed model, and production-ready artifacts that could be packaged for inference immediately after training. The client's AI development cycle shortened from weeks (infrastructure setup + trial-and-error) to days (structured runs with pre-validated configurations).