Seth/Mortdecai
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
50 Commits 1 Branch 0 Tags
0f043384e5aa8f1a75b3923bea8a77202bbe48ae
Commit Graph

4 Commits

Author SHA1 Message Date
Seth 027b835286 Session final: bakeoff fix, branding fonts, 3-GPU parallel self-play 
Current running state:
- Prod: mortdecai-v4 on RTX 4000, single-call, error correction, fall protection
- Dev: Gemini 3.1 Flash Lite (preview) + 5 bots generating training data
- Bake-off: v4 running on steel141 (3090 Ti)
- Self-play: ready for overnight — 3 GPUs parallel (3090 Ti + 2080 Ti + RTX 4000)

Changes:
- Bakeoff parser: strips think blocks, handles dict/list types
- Branding fonts: Rajdhani-Bold (official), Exo2, Orbitron, Oxanium, SpaceGrotesk
- Gemini 3.1 pricing added to cost tracker

Active data collection:
- Gemini 3.1 Flash Lite bots on dev ($20 budget, ~$4/hr with 5 bots)
- Self-play overnight: 3 tiers × 3 GPUs = ~9x throughput
- Training audit logging on all servers

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-20 00:56:45 -04:00
Seth 9d789d2524 Three-tier constraint model, mode-aware eval, boundary examples, playtest tooling 
Eval harness:
- Mode-aware scoring: sudo=strict (exact match), pray/god=soft (category match,
  in-character, appropriate intensity)
- New metrics: cmd_category_match, appropriate_intensity, scoring_mode breakdown
- Eval defaults to steel141 (192.168.0.141) — prod GPU reserved for serving

Dataset (213 examples):
- Added 31 boundary/adversarial examples (safety edges, abstention, near-boundary)
- Updated pray example reasoning: character-driven logic, not prescriptive outputs
- Tagged pray examples with scoring_mode=soft

Playtest tooling:
- whitelist.sh: add/remove/list across all 3 servers
- FRIENDS_INVITE.md + Discord version: playtester recruitment docs
- Server addresses and implementation details for both training servers

PLAN.md:
- Three-tier constraint model documented (sudo/pray/god_system)
- Success criteria split by scoring mode
- All session decisions logged

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-18 15:57:01 -04:00
Seth 48b627d498 Add LoRA training scripts and fix bake-off token budget 
- training/scripts/train_lora.py: Unsloth QLoRA trainer for qwen3:8b
- training/scripts/train_lora.sh: Launch script for steel141 RTX 3090 Ti
- eval/bakeoff.py: Fixed token budget (400->1500) that caused qwen3
  models to exhaust tokens on thinking, added --no-think flag
- agent/serve.py: Default model changed to gemma3n:e4b

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-18 10:40:18 -04:00
Seth 7da28c8800 Add model bake-off harness and base model research 
Bake-off tested 7 models on 31 seed examples via GPU-accelerated Ollama
on node-197 RTX 4000. gemma3n:e4b leads for serving (80.6% cmd match,
100% safety, 5.9s). qwen3:8b recommended as fine-tuning base (Apache 2.0,
best syntax quality, strong ecosystem). Full research in MODEL_RESEARCH.md.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-18 08:54:11 -04:00