Seth/Mortdecai
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

v3 model trained (1,308 examples, loss 0.55), API cascade, context update

Browse Source 
v3 training:
- 1,308 examples: curated + Claude-distilled + bot audit + recipes + command ref
- 1 epoch, rank 16, LR 1e-4, loss 0.55 (sweet spot)
- GGUF Q4_K_M exported, loaded in Ollama as qwen3-8b-mc-lora-v3
- Correct commands, no Chinese, proper safety refusals, dramatic God persona

API cascade for dev server:
- Stage 1: Claude Haiku ($20 budget, ~$11 spent)
- Stage 2: Gemini 2.5 Flash Lite ($20 budget)
- Stage 3: qwen3-8b-mc-lora-v3 (free, local)
- Gemini call function with persistent cost tracking
- Full status report printed at each $1 milestone

Data collection: 2,677 dev audit entries and growing
Bot status printer budget display fix

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Seth Freiberg
2026-03-19 04:52:04 -04:00
parent 234f2722db
commit e780aef8c6
5 changed files with 1130 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
PLAN.md
+4
View File
@@ -370,6 +370,10 @@ These are ideas to explore after the core system is working. Prioritize based on
| 2026-03-18 | Minecraft knowledge corpus baked into training | 1505 items, 886 recipes, 1166 blocks from minecraft-data 1.21.11. Recipe dependency trees, smelting knowledge, crafting chain examples. 107 command ref + 176 recipe examples. |
| 2026-03-18 | Claude distillation: God Soul + Haiku | God Soul document adapted from Claude's soul framework. Haiku distills 344 training examples ($0.65). Dev server switched to Haiku API ($5 budget) for high-quality live data. |
| 2026-03-18 | Version-aware training | Model trained to know it targets 1.21.x, understands 1.20.5 syntax changes, knows recipes evolve with updates. |
| 2026-03-19 | v3 LoRA training: 1,308 examples, loss 0.55 | 5.6x more data than v1. Includes Claude-distilled outputs, recipe knowledge, command reference, risk_level classification. Dramatically better than v1/v2 — correct commands, no Chinese, proper safety refusals. |
| 2026-03-19 | API cascade: Haiku ($20) → Gemini ($20) → v3 local | Dev server auto-cascades through providers as budgets exhaust. Total $40 API training data before falling back to free local model. Gemini Flash Lite validated as viable alternative. |
| 2026-03-19 | Self-service whitelist at minecraft.sethpc.xyz | Sethian Dark themed web app on CT 644, Caddy reverse proxy on CT 600. Invite key gated, whitelists on all 3 servers, only shows AI server addresses. |
| 2026-03-19 | Risk_level in model output | Model outputs risk classification (0-4) before generating commands. Validator can sanity-check: risk 0-1 should have empty commands. |
---