Highlight 8GB local training path

README.md

@@ -20,6 +20,12 @@ Alloy is an experimental 311.9M-parameter English causal language model with a 2
 
 This is a transparent small-model training project. Rather than treating data as a fixed black box, we trained in stages, evaluated each stage, inspected failure modes such as repetition and corrupted/non-English text, and revised the next corpus accordingly.
 
+## A starter-friendly local-training experiment
+
+The main value of Alloy is not frontier benchmark performance. It is a documented example of training a self-owned small language model on accessible hardware: the 312M-parameter, 2K-context Alloy stage was trained and evaluated locally on a consumer **NVIDIA RTX 3060 Ti with 8 GB VRAM**. BF16 mixed precision, microbatch size 1, gradient accumulation, and gradient checkpointing made that configuration practical.
+
+This is intended to help others quickly try a complete model-training loop on a non-datacenter computer: prepare inspectable data, train with resumable checkpoints, watch validation and hardware metrics, evaluate a frozen checkpoint, and publish reproducible artifacts. It is deliberately a starter project, not a recipe for matching models trained on large multi-GPU clusters.
+
 1. **Bronze, Silver, Gold**: three 1B-token stages established a shared 32,768-token byte-level BPE tokenizer and a 2K-context decoder-only model.
 2. **Lead / Alloy**: a 2B-token continuation used a pinned public-source mix: FineWeb-Edu (`CC-MAIN-2025-26`), English Wikipedia (`20231101.en`), and TinyStories. FineWeb-Edu documents had upstream quality score >= 3; all sources then passed English-probability, Unicode/corruption, boilerplate, URL/fragment, and repeated-text filters. Source-shard holdouts separated training from validation and test. Line breaks were deliberately preserved.
 3. The completed Alloy checkpoint was selected by its independent matching validation split, then evaluated zero-shot without prompt tuning.