CoT-Optimized-QKV 0.1.0

Chain-of-Thought Optimized QKV Attention - Infini-Attention transformers with DeepSeek reasoner CoT optimization

M3 Ultra Display Laptop?? #pip

Estimated Training Time on Apple Silicon MPS: ~2h 37m

Details:

• 22,325 training examples
• 11.3M parameter model (infini-small)
• 15 epochs, batch size 8
• 2,791 batches per epoch (41,865 total)
• 224 ms per batch
• 5,455 tokens/second

The script is ready at train_v001.py. To run full training: python3 train_v001.py

Or with fewer epochs to speed it up: python3 train_v001.py --epochs 5 # ~52 minutes python3 train_v001.py --epochs 10 # ~1h 44m

mini - Agentic training pipeline controller

mini is a CLI that uses gpt-5.1-codex-mini to orchestrate the erosolar training pipeline, manage CoT training data quality, and handle deployment checks. It is implemented in mini_the_agentic_cli.py and drives local scripts plus API calls.

What mini does

• Runs the full pipeline by invoking pipeline.py with a target master scalar and optional deploy.
• Generates training data via generate_all_training_data.py (deprecated path) and trains the model via train.py.
• Manages training samples in data_store/generated_training_data.jsonl (add/edit/delete/view with CoT diffs).
• Embeds all CoT samples with text-embedding-3-small, computes the similarity "master scalar", and analyzes or optimizes CoT consistency.
• Runs the "loser pickup" loop to add friends or update low-similarity samples using gpt-5.1-codex-mini.
• Reports attention stats and last master scalar from data_store/generation_checkpoint.json.
• Maintains conversation context with auto-squeeze and hot-swap between mini, 5.2, and 5.2-pro while preserving history.
• Verifies deployments, tests API endpoints, runs system checks, and rebuilds/deploys to Cloud Run and Firebase.
• Provides optional Tavily web search and general read/write/command tools for the agent.

Quick start

1. Set your OpenAI API key.
2. Start mini.

export OPENAI_API_KEY=...
python mini_the_agentic_cli.py

You can also store the key with --key or /key.

Wikipedia-style knowledge generator (titles only)

To create long, Wikipedia-style knowledge pairs without scraping article text:

python generate_wikipedia_knowledge.py --target 100 --resume

State is tracked in data_store/wiki_api_state.json, processed titles are logged in data_store/wiki_titles_seen.jsonl, and a persistent skip index lives at data_store/wiki_titles_seen.db. Output appends to data_store/generated_training_data.jsonl.

mini shortcuts:

• /wiki runs the generator with --target -1 (all titles, resumable).
• MINI_WIKI_TARGET overrides how many records mini generates per run (0 disables; <=0 means all).

Coding-only generator

To create a coding-only slice of training data each run:

python generate_coding_only.py --target 200 --resume

State is tracked in data_store/coding_only_state.json, seen prompts in data_store/coding_only_seen.db, and prompt logs in data_store/coding_only_prompts.jsonl.

mini shortcuts:

• /coding runs the generator with --target -1 (all prompts, resumable).
• MINI_CODING_TARGET overrides how many records mini generates per run (0 disables; <=0 means all).

Startup behavior

• Runs a quick self-test unless --no-self-test is set.
• Checks status on launch.
• If models/erosolar is missing, runs the pipeline once (no deploy).
• If Cloud Run or Firebase is not marked as deployed in version.json, triggers auto-deploy.

CLI shortcuts

python mini_the_agentic_cli.py --run
python mini_the_agentic_cli.py --status
python mini_the_agentic_cli.py --self-test
python mini_the_agentic_cli.py "generate 1000 records"
python mini_the_agentic_cli.py --no-auto
python mini_the_agentic_cli.py --no-self-test

Interactive commands

/help /status /run /generate /train /deploy /version /key /exit

You can also type natural language requests; mini decides which tools to run.

Data and state

mini reads and writes:

• data_store/generated_training_data.jsonl (training samples)
• data_store/cot_embeddings.json (CoT embeddings)
• data_store/version.json (data stats)
• data_store/generation_checkpoint.json (master scalar stats)
• models/erosolar (trained model check)
• version.json (erosolar version and deployment status)
• ~/.agi/secrets.json (stored API key)
• ~/.agi/mini_context.json (conversation context)

Dependencies and external tools

mini uses:

• OpenAI Responses API and Embeddings API (requires OPENAI_API_KEY)
• Python packages: aiohttp, httpx, numpy (API calls and embeddings) and rich (optional UI)
• gcloud and firebase CLIs for deployment commands