80 blogs is enough for fine-tuning a small model on your brand voice, but the real bottleneck won’t be the generation — it’ll be maintaining quality at scale. A few practical steps:
Fine-tune a small LLM (Mistral 7B or Llama 3 via LoRA) on your 80 blogs as context + your existing social posts as target outputs. Focus on tone, structure, length constraints. Prompt engineering first — before training, try few-shot prompting with your best 5-10 blog-to-social examples. That alone might get you 80% of the way. Build a review loop — generate 5 variants per post, have a human pick the best. Use those picks as your training set for iteration.
Training on 80 samples won’t produce production-ready output on its own, but as a bootstrap for a human-in-the-loop process it works fine. Scale the dataset by keeping every approved generation.