melissa kate
<< Little Jane

tiny-prose running cost ledger

Rough tracking of cloud spend across the project. Update as we go.

dateruncostnotes
2026-06-30v1 pretraining (33M / 11M tokens)~$0.80first pretraining attempt
2026-06-30v2 pretraining (33M / 16M tokens)~$1.00corpus expansion round 1
2026-07-01v3 pretraining (100M / 395M tokens)~$5.60BL Books added, big jump
2026-07-01wasted runs (aborted tokenizer / stuck encoding)~$2-3pre-cache-invalidation, before figuring out fix
Modal total~$10-15credits depleted; migrated to Lambda

Lambda (GH200 $2.29/hr, H100 SXM5 $4.29/hr, H100 PCIe $3.29/hr)

Billing switched to per-minute as of 2026-08-19 (was per-hour).

dateruninstancedurationcostnotes
2026-07-02v4 pretraining (200M / 391M tokens)H100 SXM52.25 hr$9.63includes OOM restart + tokenizer + encoding + train + eval
2026-07-03Phase 2 SFT (chat fine-tune, 51K pairs)H100 PCIe0.67 hr$2.19best val 2.044
2026-07-03Phase 3 candidate generation (300 prompts × 5)H100 PCIe0.14 hr$0.47for rejection-sampling curation
2026-07-05v5 pretraining (200M / 270M words)H100 PCIe1.32 hr$4.34full OCR/inline-marker cleanup pass; final val 3.113 (new tokenizer, not comparable to v4)
2026-08-19v11 pretraining (245M / 4.92B tokens)GH200 480GB6.48 hr$16.8618L/16H/1024, block 2048, batch 48, bf16 autocast + tf32 + torch.compile, 211K tok/s avg; final val 2.7814, no overfit across 4.9 epochs; figure covers benchmark sweep, a relaunch after SSH key loss, and transfer time
2026-08-20v11 anneal sweep (4 variants)GH200 480GB0.7 hr$1.56plays-vs-books mix sweep; 4 runs x 3.8 min at 217K tok/s; anneal WORKS — replies drop to 1-2 sentences. (August total $18.42 - $16.86 pretrain.)
Lambda total$35.05

Grand total so far: ~$45-50

Upcoming (rough estimates)