Gollemer

Gollemer is a high-performance Mixture of Experts (MoE) neural network framework and training pipeline written entirely in Go. It is designed for maximum performance with zero external dependencies, featuring a native core optimized for SIMD-accelerated CPU training and an autonomous adaptive supervisor.

---

⚡ Quick Start

1. Installation

Gollemer requires Go 1.26+ to leverage native SIMD acceleration.

git clone https://github.com/golangast/gollemer
cd gollemer
export GOEXPERIMENT=simd
go mod tidy

2. Training the Model

We provide a simplified Makefile to handle the curriculum training process.

make train
make train ARGS='-cartridges="data/models/intents/computer.cartridge" -data "data/training/trainingdata/computer/computer.csv"'

3. Running the LLM

Chat with your trained model using the interactive shell.

make chat
# or use 'make llm'

---

🛠️ Makefile Commands

Command	Description
make train	Start fresh curriculum training (clears MoE models, preserves word2vec)
make train-fresh	Full cold start — clears ALL models including word2vec, then trains
make train-small	Run small social dataset, print loss + memory, and test the model
make train-small-seq2seq	Run strict pure Q→A seq2seq tiny demo
make test-small-seq2seq	Load tiny seq2seq model and probe a few prompts
make seq2seq-prompt PROMPT="hello"	Send a custom prompt to the saved tiny seq2seq model
make seq2seq-chat	Start an interactive tiny seq2seq chat loop
make metrics	Run metrics aggregation and CSV export for edit logs
make export-labels	Export training examples to CSV for manual labeling
make install-hooks	Install Gollemer Git pre-commit validation hook
make clean	Remove MoE model checkpoints (preserves word2vec)
make clean-all	Remove ALL model files including word2vec
make help	Display available commands
make conversing-pb	Convert conversing.yaml → conversing.pb (protobuf dataset)

Usage example with custom data:

make train ARGS='-cartridges="data/models/intents/computer.cartridge" -data "data/training/trainingdata/computer/computer.csv"'

---

Small Training (make train-small)

make train-small runs a compact social-curriculum training loop on the tiny demo dataset (small_social_demo.pb or .csv). It:

1. Trains a small MoE model using a direct answer-only objective to force rapid loss descent.
2. Prints heap and system memory stats before and after training.
3. Probes the trained model with four fixed prompts (hello, what is your name, how are you, can you help me) and prints the generated responses plus latency.

This is useful for quickly verifying that training, inference, and the model loader work end-to-end without running the full curriculum.

---

Protobuf Datasets

Training data can now be loaded from protobuf (*.pb) files in addition to CSV. The small-training and seq2seq pipelines prefer .pb datasets when available:

• data/training/trainingdata/conversing.pb — multi-turn conversations (converted from YAML via make conversing-pb)
• data/training/trainingdata/small_social_demo.pb — tiny social demo dataset

The proto definitions live under internal/ai/training/proto/ and internal/ai/training/proto/dataset/.

---