Skip to content

MODEL_CARD.md

Latest commit

 

History

History
151 lines (119 loc) · 5.84 KB

MODEL_CARD.md

File metadata and controls

151 lines (119 loc) · 5.84 KB
Raw

Aura Model Card

Lane defaults below are core/config.py:LLMConfig (verified 2026-08-01). Override any of them with AURA_MODEL, AURA_DEEP_MODEL, AURA_BRAINSTEM_MODEL, AURA_FALLBACK_MODEL, or the nested AURA_LLM__* form.

Primary Model (Cortex)

Field Value
Role Primary reasoning and conversation
Architecture Transformer LLM (32B parameters, Qwen2.5-32B-Instruct)
Runtime MLX on Apple Silicon
Quantization 8-bit (MLX native); a 4-bit build exists as a legacy / low-memory option
Context Window 8192 tokens (configurable)
Inference Local, on-device
Fine-tuning 8-bit base weights with a personality LoRA applied; the runtime can also promote a fused LoRA delta (training/fused-model/active.json) as the live Cortex without a re-quantize

Intended Use

Primary model for all user-facing conversation, reasoning, tool planning, and complex cognitive tasks.

Limitations

  • May confabulate when knowledge is insufficient
  • Context window limits multi-turn reasoning depth
  • 4-bit quantization trades precision for memory efficiency
  • Cannot process images or audio natively

Ethical Considerations

  • Model weights are publicly available base models
  • No private/personal data in training
  • Prompt injection mitigations applied at runtime layer

Deep Model (Solver)

Field Value
Role Deep-reasoning hot-swap tier for hard problems
Architecture Transformer LLM (72B parameters, Qwen2.5-72B-Instruct)
Runtime MLX on Apple Silicon
Quantization 4-bit (MLX native)
Inference Local, on-device

Intended Use

Hot-swapped in for the deepest reasoning passes on 64GB-class desktops. It is the highest-capacity local lane but the slowest (~84s/pass), so it is not the default foreground model — the 32B Cortex handles standard turns and the Solver is promoted only when a problem warrants it. Auto-detected/enabled via AURA_DEEP_MODEL.

Background Model (Brainstem)

Field Value
Role Background maintenance, classification, lightweight tasks
Architecture Transformer LLM (7B parameters)
Runtime MLX on Apple Silicon
Quantization 4-bit (MLX native)
Context Window 4096 tokens
Inference Local, on-device

Intended Use

Background tasks: memory consolidation, classification, health probes, maintenance reasoning. Never used for user-facing responses in production mode.

Limitations

  • Reduced reasoning capability compared to primary
  • Not suitable for complex multi-step reasoning
  • Background-only; foreground lane isolation prevents interference

Reflex Model

Field Value
Role Fast reflex lane: sub-second acknowledgements, routing, guards
Architecture Transformer LLM (1.5B parameters, Qwen2.5-1.5B-Instruct)
Runtime MLX on Apple Silicon
Quantization 4-bit (MLX native)
Inference Local, on-device

Intended Use

The lowest-latency local tier. Handles reflexive turns and lightweight routing/guard decisions when the 32B Cortex is warming or contended, so the conversation lane can answer immediately instead of waiting on the heavy model. Never used for substantive full-mind replies.

Cloud Fallback Model

Field Value
Role Fallback when local models unavailable
Provider Google Gemini (llm.teacher_model, default gemini-2.5-pro) via core/brain/llm/gemini_adapter.py
Credential GEMINI_API_KEY / llm.gemini_api_key — absent by default
Activation Opt-in only; requires explicit configuration
Privacy Prompts classified before transmission

Intended Use

Emergency fallback only. Activated when local models are unavailable and the user has explicitly opted in. With no API key configured — the default state — this lane cannot activate at all.

Privacy Controls

  • Cloud use is a governed action domain (ActionDomain.CLOUD_FALLBACK in core/governance/will.py), so it passes the Unified Will and leaves a receipt like any other consequential action.
  • Prompts are classified before transmission; sensitive and restricted content is not sent.
  • Cloud usage is logged in the Will receipt trail and is visible in the audit chain export.

Model Verification

Model identity is measured from the artifact, not inferred from its path. core/brain/llm/model_artifact_profile.py exists because footprint, minimum-headroom, deadline, cache-residency, and identity decisions used to be derived from spoofable path substrings ("72b", "cortex", "zenith") — a renamed heavy checkpoint inherited light-model budgets, and an unrelated path containing "32b" inherited a 20 GB reservation.

At load time the runtime reads:

  • model.safetensors.index.json → metadata.total_parameters and metadata.total_size (exact weight bytes),
  • config.json → architecture shape (parameter count is estimated from this when index metadata is absent),
  • the safetensors file listing (names + sizes).

From those it derives a cached ModelArtifactProfile carrying a SHA-256 fingerprint over config bytes, index metadata, and the weight-file listing. This is an identity binding, not a weight hash — hashing 20 GB on every admission check is not viable, and the card should not imply otherwise. The fingerprint changes whenever the artifact's declared shape changes.

The profile records which evidence produced it, so a receipt can distinguish measured truth from a naming-convention fallback (used only when the artifact is absent, as in tests and pre-download paths).

Memory footprint is then validated against the hardware profile before the load is admitted; AURA_MLX_32B_LOAD_MIN_AVAILABLE_GB sets the floor below which a 32B load is refused outright.