Scale training budget — full example (illustrative USD)
Written by Grok - Human checking required
Training-lead voice. Numbers are order-of-magnitude scenarios for planning discipline, not vendor quotes.
Hard gate T8 until capital exists. Nano laptop costs are published separately and must not be laundered into this table.
1. Objective of first serious public epoch
| Field | Example target (human may retarget) |
|---|---|
| Ambition class | Public model with full transparency stack; trajectory toward OLMo/LLM360 class of openness, not a false claim of current parity |
| First milestone | e.g. 7B-class dense public LM with dense or schedule-defined public checkpoints |
| Non-goals | Closed fine-tune marketed as TTLLM without lineage; paywalled verify |
| Success metric | Third party re-hashes manifests; ttlink over published corpus; stream complete; eval honesty pack with explicit non-claims |
2. Preconditions (gates and bone)
| Precondition | Status type | Notes |
|---|---|---|
| BOUNDARY unchanged | Process | Domain 1 if any movement |
| Entity for contracts/keys | Hard T6 | Formation pack exists |
| Production key policy | Hard T5 | Ceremony example exists |
| Data mixture card | Writable bone | ttllm.data_mixture.v1 spec |
| Checkpoint publication policy | Writable bone | Interval frozen before train |
| R2 / object store for weights | Hard T2 until enable | Design pack exists |
| Standing adversarial capacity | Hard T10 hire or interim harness-only honesty | Harness is not a person |
3. Budget sketch (illustrative ranges)
| Line | Low (USD) | High (USD) | What it buys | Transparency constraint |
|---|---|---|---|---|
| GPU hours (train + ablations) | 80,000 | 400,000 | Cluster time | Public train code + hyperparams |
| Storage + egress (public artefacts) | 5,000 | 40,000 | Checkpoints, data docs, indexes | Public hashes; no paywall on verify |
| Data licensing / curation / legal review | 10,000 | 100,000 | Lawful mixture | Domain 3 process; DATA_CARD |
| Engineering labor (6–12 mo) | 150,000 | 400,000 | Train + infra + seals | Influence disclosure if merge/keys |
| Standing red team / campaigns | 40,000 | 120,000 | Domain 10 | Default publish findings |
| Eval / contamination work | 15,000 | 60,000 | Claim gate compliance | Honesty packs |
| Legal / entity / covenant execution | 15,000 | 50,000 | T6 close path | Public non-sensitive note |
| Contingency (15%) | ~45,000 | ~175,000 | Overrun | Same constraints |
| Total (approx.) | ~360,000 | ~1,350,000 | First serious epoch | Free core stays free |
Human must replace ranges with quotes before treating as a board budget.
4. Unit economics link
- Nano actuals:
docs/business/NANO_UNIT_ECONOMICS_ACTUAL.md+scripts/nano_cost_ledger.py - Nano wall-clock must never be cited as evidence of scale readiness
- Scale success = public artefacts strangers can verify, not press release FLOPs
5. Spend controls (ethos)
- No invoice line may fund free-core paywall tooling.
- Public checkpoint interval set in Domain 1 before train starts.
- Stream events required: data_prepared, training_started, loss/checkpoint cadence, training_finished, release, ttlink_index_sealed.
- If budget cut forces opacity, cut capability claims first — not the skeleton.
6. Explicit non-claims
This document does not authorise spend, hire, or raise. It does not assert that TTLLM has completed a 7B or 32B run.
Written by Grok - Human checking required — also on https://ttllms.com/placeholders/