egonetics-evolution

Egonetics Evolution Layer — autonomous LLM training experiments.

Philosophy

karpathy autoresearch style: single-file, agent-editable, fixed time budget, metric-driven keep/discard.

Three files:

File	Edited by	Purpose
prepare.py	Human (fixed)	Training data generation via SEAI
train.py	Agent	Model architecture + training loop (GPT + MoE)
mini_dsv4.py	Agent	Mini DeepSeek-V4 architecture (MLA + MoE)

Dependency

Calls SEAI for data preparation:

/Users/Shared/SubjectiveEgoneticsAI/

Python 环境（两个 venv，不要混）

脚本	环境	命令
train.py / mini_dsv4.py	本仓 .venv（Python 3.11 + torch）	source .venv/bin/activate 后用 python …
prepare.py	~/llama-factory/venv（SEAI 同款）	./scripts/run-prepare.sh --exp-id <id>

首次建训练环境：

cd /Users/Shared/egonetics-evolution
python3.11 -m venv .venv && .venv/bin/pip install "torch>=2.5"

Cursor 解释器选 egonetics-evolution/.venv。详见 project_config.md。

Scale Presets (mini_dsv4.py)

source .venv/bin/activate
python mini_dsv4.py --scale tiny    # ~25M  params, quick test
python mini_dsv4.py --scale small   # ~150M params, trainable locally
python mini_dsv4.py --scale medium  # ~500M params
python mini_dsv4.py --scale base    # ~850M params, vs Qwen2.5-0.8B

Experiment Runner

Driven by TypeScript experiment-runner.ts via seai-bridge.ts:

1. ./scripts/run-prepare.sh --exp-id <id> — generates training data (SEAI)
2. train.py --exp-id <id> or mini_dsv4.py --exp-id <id> — runs training (.venv)
3. Parses val_loss from stdout JSON
4. Keeps or discards based on metric improvement

Ecosystem runtime CFL (2026-05-18)

The PRD-driven Evolution ecosystem runtime lives in ecosystem-runtime/. It is a Go runtime that reuses the standalone P/R/V/S CFL projects and preserves the corrected 2026-05-17/18 PRD semantics: open tasks, non-exclusive claims, independent seed runtimes, free communication, V-owned evaluation dimensions, and user resource injection rather than user payment.

See CHRONICLE.md and ecosystem-runtime/README.md for the Chronicle ↔ Git handoff record.