UNNAMED AI

Assembling… Integration 19%

Data pipeline

dedupe · filtering · licensing

Tokenizer

BPE/Unigram

Pretraining

Causal LM (stage A)

SFT

instruction tuning

RLHF / DPO

reward model

Evals

MMLU · MT-Bench

Safety

red teaming · filters

Serving

vLLM/TensorRT-LLM

RAG & Memory

vector index

Training metrics (simulated)

global step

—

tokens / s

—

loss

—

throughput (est.)

—

shard

3 / 64

checkpoint

—

Cluster (synthetic)

nodes × gpus

4 × 8

avg GPU util

—

VRAM / GPU

80 GB

job status

pretrain (A)

Artifacts (local hashes)

dataset.manifest.json

hashing…

tokenizer.json

hashing…

checkpoint-0003.safetensors

hashing…

eval-report.md

hashing…

Hashes computed locally (WebCrypto or fallback) for reproducibility.

We’ll email you once UNNAMED AI is available.