Sonapse-LAB / docs /DATA_CARD.md
Kai9987kai's picture
Sync GitHub 6f2761b with Hugging Face LFS assets
713d4ed
|
Raw History Blame Contribute Delete
10.2 kB
# SYNAPSE Data Card
## Dataset & Environment Specifications
### 1.5 Landmark Biological Drosophila Connectome Datasets
The architecture supports real-time multi-connectome grafting across eight landmark community datasets:
1. **MaleCNS v1.0**: HHMI Janelia FlyEM / Cambridge / MRC LMB / Google Research (2026-06-08, CC-BY-4.0, 269 projects). Whole CNS (brain + VNC).
2. **BANC (Brain And Nerve Cord)**: Harvard Medical School (Lee & Wilson labs) / Princeton (Murthy lab) / Harvard (Drugowitsch lab) (2026, CC-BY-4.0, 22 projects).
3. **FlyEM Optic Lobe Connectome**: HHMI Janelia FlyEM / Google Connectomics (2024, CC-BY-4.0, 5 projects). Retinotopic visual columns.
4. **FANC (Female Adult Nerve Cord)**: Harvard Medical School (Wei-Chung Allen Lee lab) (2024, unknown, 3 projects).
5. **FlyWire (FAFB v783)**: Princeton University (Murthy & Seung labs) with FlyWire consortium (2024, CC-BY-NC-4.0, 180 projects). Whole adult female brain (139,255 neurons).
6. **MANC (Male Adult Nerve Cord)**: HHMI Janelia FlyEM / Cambridge / Google Research (2023-06, CC-BY-4.0, 17 projects).
7. **Larval Drosophila Brain Connectome**: University of Cambridge (Michael Winding et al.) (2023-03, CC-BY-4.0, 7 projects). Complete insect larva (3,016 neurons).
8. **Hemibrain**: HHMI Janelia Research Campus (FlyEM Project Team) (2020-01-22, CC-BY-4.0, 23 projects). Central brain hemisphere.
### 1. Embodied Simulation Worlds
SYNAPSE supports both the 9 native GraftBrain simulation tasks and 4 advanced frontier environments implemented in `ComplexWorld`:
#### Native Tasks
1. `forage`: Locate and consume food targets while avoiding boundary hazards and managing energy decay.
2. `maze`: Multi-corridor navigation requiring working memory and wall avoidance.
3. `push`: Manipulate external objects into designated target zones.
4. `grasp_place`: Position arm endpoint, engage grip, and transport objects.
5. `peg_insert`: High-precision 3D alignment and insertion task.
6. `delayed_match`: Encode color/spatial cue, navigate through an occlusion delay, and select matching resource.
7. `go_nogo`: Discriminative action selection conditioned on cue frequency/polarity.
8. `button_sequence`: Sequential activation of multi-state targets in ordered permutations.
9. `webcam_chase`: Track an orb that can be driven by webcam motion in the upstream environment.
#### Frontier Complex Ecology Tasks (`ComplexWorld`)
10. `ecosystem`: Comprehensive virtual biosphere combining autonomous moving predator drones, variable terrain friction patches (ice, mud, sand), food targets, dynamic gates, and day/night illumination cycles.
11. `predator_evasion`: High-threat evasion arena where the agent must out-maneuver pursuit drones, conserve stamina, and utilize terrain friction to escape predation.
12. `dynamic_arena`: Multi-room navigational maze featuring harmonic oscillating gates and variable friction surfaces testing spatial mapping under traction variations.
13. `puzzle_chamber`: Multi-stage logic environment where the agent must press interactive floor triggers/buttons to open pressure-locked security gates protecting high-yield food caches.
#### Deep Research 9-Arm Connectome Control Topologies (`synapse.connectome_controls`)
1. `bio`: True biological Janelia Drosophila male-CNS subcircuit with signed Dale synapses.
2. `degree_rewired`: Maslov-Sneppen degree-preserving directed rewiring.
3. `sign_rewired`: Sign-preserving rewiring testing Dale's law constraints.
4. `community_rewired`: 4-cluster modular community-preserving rewiring.
5. `random_sparse`: Erdős-Rényi random sparse graph with identical edge count and density.
6. `learned_sparse`: Learned sparse weight matrix with equal compute/parameter budget.
7. `unsigned`: Real graph with neurotransmitter signs removed.
8. `local_only`: Real graph with long-range edges pruned.
9. `brain_cord_motif`: Local sensory-motor loops with central supervisory hub (2026 brain-cord abstraction).
#### Deep Research Machine-Verifiable Preference Corpus (`synapse.preference_dpo`)
- Automatic first-party preference pairs $(x, y_w, y_l)$ generated from machine-checkable outcomes:
- Safety: Surviving vs. fatal trajectory actions.
- Return: High-return vs. low-return planner rollouts.
- Calibration: Well-calibrated vs. overconfident forecasts.
- Code/Patch: Invariant test-passing vs. failing patch proposals.
#### SAIL Verifiable Curriculum & ReLIFT Tasks (`synapse.verified_curriculum`)
14. `verifiable_relift`: Procedurally generated sequential challenge stages with strictly verifiable intermediate subgoals:
- `CUE_DISCOVERED`: Visual recognition and approach toward target color cues.
- `GATE_TRIGGERED`: Exact physical manipulation of spatial floor switches/buttons.
- `BARRIER_PASSED`: Safe traversal through dynamic barrier passages without collision.
- `TARGET_REACHED`: Successful goal acquisition and consumption.
- Each transition and subgoal verification is cryptographically anchored by a SHA-256 execution receipt.
#### SAIL 4-Arm Continual Learning Milestone (`synapse.flagship_benchmark`)
- Standardized sequential curriculum evaluating catastrophic forgetting:
$$\text{Task}_1 (\text{forage}) \rightarrow \text{Task}_2 (\text{delayed\_match}) \rightarrow \text{Task}_3 (\text{push}) \rightarrow \text{Task}_4 (\text{ecosystem})$$
- Evaluates four comparative arms:
1. `fixed_gradient`: Standard Adam backpropagation baseline (prone to catastrophic forgetting).
2. `dynamic_plasticity`: Structural neurogenesis + Hebbian weight plasticity.
3. `continual_sleep`: Trainable student-teacher sleep consolidation.
4. `full_developmental_sail`: Full integration (Continual Backprop + Fast Memory + Sleep + MPC Planning).
#### Task-Aware Planning Study Data (14 November 2026)
This is locally generated synthetic-world evaluation data, not an external
dataset. Six fresh seeds were paired across `forage`, `delayed_match`, and
`button_sequence`; each seed/task unit shared 64 exploratory support transitions
before four frozen query arms (actor, horizon 1, horizon 3, task-aware horizon
3). Each query episode stopped at an environment terminal event or 256 steps.
Only environment-reported terminal events count as success, failure, or timeout;
episodes without one are recorded as incomplete.
| Task | Episodes per arm | Task-aware outcomes | Task-aware minus actor mean return |
|---|---:|---|---:|
| `forage` | 6 | 0 success, 0 fail, 0 timeout, 6 incomplete | −0.6848 |
| `delayed_match` | 6 | 0 success, 1 fail, 0 timeout, 5 incomplete | 0.0000 |
| `button_sequence` | 6 | 0 success, 0 fail, 0 timeout, 6 incomplete | 0.0000 |
Across 18 paired seed/task units, no arm achieved a success. Each arm recorded
one terminal failure and 17 incomplete episodes. The task-aware completion
difference was 0.000 with descriptive paired 95% interval
`[0.000, 0.000]`; mean return difference was −0.2283 (`[−0.4863, −0.0274]`).
The run used 18,832 transitions against a 32,768 limit. Most episodes were
right-censored by the query cap, so this study does not resolve terminal task
performance. The [full report](results/2026-11-14-task-aware-planning.md),
[frozen protocol](results/2026-11-14-task-aware-planning-protocol.json), and
[receipt](results/2026-11-14-task-aware-planning-receipt.json) preserve the
complete outcomes and provenance.
The confirmation did not include a separate task-context ablation, so it does
not identify the contribution of individual context inputs.
### 2. Dynamic Physics & Ecological Mechanics
- **Autonomous Predator Drones**: Drones execute continuous pursuit vectors toward the agent at speed $v_{\text{predator}} = 0.04$, generating a continuous distance-dependent threat field:
$$\text{Threat}(d) = \max\left(0, 1 - \frac{d}{R_{\text{hazard}}}\right)$$
- **Terrain Friction Zones**: Continuous rectangular patches modifying effective friction and stamina expenditure:
- *Ice*: Low friction ($0.1\times$), high slip, low stamina cost.
- *Mud*: High friction ($2.5\times$), low slip, high stamina cost.
- *Sand*: Moderate resistance ($1.8\times$).
- **Dynamic & Puzzle Gates**:
- *Harmonic Oscillating Gates*: Cycle between open (passable) and closed (solid barrier) states on fixed periodic intervals.
- *Button-Locked Gates*: Remain solid barriers until the agent depresses an associated trigger button in the environment.
- **Diurnal Ambient Illumination Cycle**: Sine-wave ambient lighting modulating visual rangefinder rays and sensory contrast:
$$I_{\text{ambient}}(t) = 0.5 + 0.4 \cos\left(\frac{2\pi t}{T_{\text{day}}}\right)$$
### 3. Observation Structure (155 continuous float32 values)
- `[0:33]`: Ray depth sensor (33 visual rangefinder rays across 120° FOV), modulated by ambient light and ray occlusion.
- `[33:66]`: Red channel visual contrast.
- `[66:99]`: Green channel visual contrast.
- `[99:132]`: Blue channel visual contrast.
- `[132:136]`: Normalized arm reach, arm height, grip, and whether an object is held.
- `[136:140]`: Normalized forward/angular velocity, sine and cosine of body heading.
- `[140:149]`: One-hot native/frontier task identifier.
- `[149:153]`: Environmental associative cues for Mushroom Body olfactory memory.
- `[153]`: Previous observed reward.
- `[154]`: Current time divided by episode horizon.
### 4. Connectome Topologies
- **FlyWire Profile**: Synthetic model of central Drosophila brain neuropils with 25% inhibitory GABAergic/glutamatergic fractions.
- **MANC Profile**: Ventral nerve cord motor and ascending/descending interneuron clusters.
- **Optic Lobe Profile**: Visual projection columns with retinotopic receptive fields.
- **Janelia Male CNS v1.0 (Optional External Data)**: Full male fly CNS connectome (11,751 cell types, 3,830,931 type-to-type edges, CC BY 4.0), loaded from `malecns_types.npz`.
### 5. RSI Benchmarks & Task Partitions
- Synthetic feature evaluation splits:
- Development families `0-1`: Initial policy search and hyperparameter calibration.
- Validation families `2-3`: Intermediate meta-selection and mutation screening.
- Sealed families `4-5`: Strictly isolated query tasks evaluated post-adaptation with external writer disabled.