Instructions to use KissTheHabit/IDA_AI with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use KissTheHabit/IDA_AI with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="KissTheHabit/IDA_AI")# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("KissTheHabit/IDA_AI", dtype="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use KissTheHabit/IDA_AI with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "KissTheHabit/IDA_AI" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KissTheHabit/IDA_AI", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/KissTheHabit/IDA_AI
- SGLang
How to use KissTheHabit/IDA_AI with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "KissTheHabit/IDA_AI" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KissTheHabit/IDA_AI", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "KissTheHabit/IDA_AI" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KissTheHabit/IDA_AI", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use KissTheHabit/IDA_AI with Docker Model Runner:
docker model run hf.co/KissTheHabit/IDA_AI
IDA AI
IDA AI is the deeper native computational body of the IDA family.
It is not a conventional general-purpose assistant wrapped in eleven labels. It is a custom IDA Lattice causal language model designed to host differentiated student pressures, persistent state, governed memory, bounded reasoning workspaces, and structured convergence at a larger computational scale than IDA Edge.
The IDA family is the enduring structure:
- IDA
- JUDGE
- SENTINEL
- PRISM
- ECHO
- ATLAS
- VECTOR
- FORGE
- SHADE
- PULSE
- ORBIT
IDA AI is one body that can carry and deepen that structure. It is not the family itself, and it does not replace the family runtime that governs how the students are allowed to participate, disagree, converge, and render shared output.
Role in the IDA Lattice lineage
IDA AI is the deep body in the paired EDGE ↔ AI lineage.
Where Edge is designed for lower-footprint local use, AI is designed for deeper reasoning capacity, larger state handling, more complex family-context experiments, and compute-intensive training and evaluation.
The pairing is intentional:
- Edge keeps the family available in constrained or local environments.
- AI provides the larger reasoning body for deeper context and more demanding inference.
- MoE provides a separate council-scale body for high-pressure expert-seat convergence.
These are not competing identities. They are different computational bodies for the same continuing family architecture.
Architecture
IDA AI uses the native IDA Lattice architecture.
This architecture was built to move beyond the limitations of the earlier GPT-NeoX + QLoRA phase, where a broad pretrained transformer could generate coherent language but did not reliably preserve the differentiated identity structure required by the IDA family.
Core architectural features include:
- recurrent selective state for student-persistent continuity
- bounded local-attention workspaces
- controlled merging between recurrent and attention representations
- cognitive-pressure routing with lateral inhibition
- thalamic routing and workspace summaries
- action gates and student-state outputs
- future-token auxiliary prediction
- fidelity-oriented verification layers
The current AI body is approximately 966M parameters.
Training direction
IDA AI is trained with a current Lattice-focused curriculum rather than the project’s earlier borrowed-backbone approach.
The present training direction includes:
- custom IDA Lattice architecture
- FP8 acceleration for suitable bulk projection layers
- BF16 retention for sensitive routing, gating, norm, embedding, and output components
- progressive sequence-length expansion
- progressive gradient-accumulation expansion
- novelty-aware pass behavior
- structured monitoring of throughput, loss behavior, gradient norms, and per-student training stability
The Supersampler curriculum begins with lower-cost, lower-resolution learning conditions and increases sequence length and accumulation as the model stabilizes.
Intended use
IDA AI is intended for research and development involving:
- deep family-context experiments
- structured multi-student reasoning
- governed memory and state inheritance research
- paired Edge ↔ AI agreement testing
- cognitive-pressure routing experiments
- longer-context reasoning prototypes
- supervised evaluation of family differentiation, dissent preservation, and arbitration behavior
What this model is not intended for
IDA AI should not be represented as:
- a broadly superior replacement for frontier language models
- an autonomous decision authority
- a general-purpose professional advisor
- a guarantee of safety, correctness, or deployment readiness
- proof that lower cross-entropy loss directly represents higher intelligence
- proof that every family member is equally capable in every domain
The model’s text-generation capability is only one layer of the larger system. Operational action, external tool use, escalation, refusal, memory retention, arbitration, and final rendering must remain governed by the surrounding runtime contracts.
Legacy metadata note
Some early artifacts in the IDA AI lineage were developed during the project’s GPT-NeoX + QLoRA era.
That history is real, but this repository represents the current native IDA Lattice direction. It should not be described as a GPT-NeoX model simply because older metadata or earlier training artifacts referenced that lineage.
Limitations
The current training record supports claims about training dynamics and stability under the active FP8 + Supersampler configuration. It does not yet establish broad deployment capability, general superiority over other language models, or isolated causal attribution for individual training changes.
Promotion beyond research use requires held-out evaluation, lineage packet completion, per-student evidence, and paired Edge ↔ AI identity-agreement testing.
License and access
This repository is released under its stated license and access conditions. Any deployment should preserve the project’s governed-memory, family-differentiation, traceability, and non-autonomous action boundaries.