fine-tuning DAG — one YAML reproduces everything
data
Datatrove dedup
Nemotron-CC + Presidio
contamination
MinHashLSH vs
MMLU-Pro / MT-Bench
SFT
Axolotl v0.8 + ZeRO-3
2-3 epochs 8xH100
DPO / GRPO
TRL 0.15
1 epoch, beta sweep
quantize
GPTQ + AWQ + GGUF
Marlin INT4
serve
vLLM 0.7 + EAGLE-3
K8s HPA on queue-wait
eval
lm-eval + MT-Bench v2
RewardBench-2 + MMLU-Pro
safety eval + model card
Llama Guard 4 + ShieldGemma-2 + MOF 2026 template
content-hashed manifest + reproducibility contract
every stage consumes and produces a content-hashed artifact
re-running with the same YAML + seed yields identical hashes end to end
one command: ./pipeline.sh config/llama3.3-8b-domainX.yaml
Langfuse traces training+inference; W&B run links in model card
ablation table baked in: base / SFT-only / SFT+DPO / SFT+GRPO
serving target: 2.5x+ throughput vs baseline with EAGLE-3; $/1M tokens reported vs hosted APIs