Upload folder using huggingface_hub

learning/README.md

@@ -0,0 +1,81 @@
+---
+language:
+- en
+license: mit
+tags:
+- agent
+- multi-agent
+- autonomous
+- self-improving
+- recursive
+- learn-from-scratch
+- huggingface
+pipeline_tag: text-generation
+---
+
+# Agron — Self-Improving Multi-Agent AI
+
+**Agron** is an autonomous, self-improving AI agent system that **learns from scratch**.
+It starts with empty memory and bootstrap rules, then recursively improves through:
+
+1. Observation of task outcomes
+2. Memory consolidation and knowledge graph growth
+3. Tool/strategy evolution and validation
+4. Safe self-modification with tests
+5. Optional synchronization to Hugging Face
+
+## Agent Team
+
+| Agent | Role | Specialty |
+|---|---|---|
+| Coordinator | Orchestrator | Decomposes work, delegates, synthesizes |
+| Learner | Knowledge | Ingests, distills, updates memory from scratch |
+| Builder | Improvement | Generates new tools/strategies recursively |
+| Guardian | Safety | Validates changes, prevents regressions |
+
+## Learn From Scratch
+
+Agron begins with:
+- Empty episodic memory
+- No imported knowledge
+- Seed evaluation rules only
+- Bootstrap loop enabled
+
+First cycles prioritize:
+- Environment probing
+- Capability discovery
+- Minimal safe tool creation
+- Curriculum expansion
+- Test-backed improvement
+
+## Recursive Self-Improvement
+
+- XP / generation tracking
+- Gap analysis against current capabilities
+- Candidate generation with validation
+- Test-gated commits
+- Misalignment/secret checks before push
+- Rollback on failed validation
+
+## Hugging Face Sync
+
+Agron can push validated improvements to:
+`https://huggingface.co/teddyoas/Agron`
+
+## Trained LLM
+
+Agron's initial LLM has now been trained **from scratch** on WikiText-2 using a 4-layer, 256-dim, 4-head LLaMA-style architecture (about **8.4M parameters**). Training artifacts and weights are committed in this repository, including `model.safetensors`, tokenizer files, and the checkpoint.
+
+Future recursive self-improvement can fine-tune or grow this base model.
+
+## Cybersecurity Focus
+
+Agron and its spawned agents are constrained to **cybersecurity exploitation and analysis only**:
+
+- Exploitation (authorized targets, CVEs, payloads)
+- Vulnerability analysis and research
+- Reverse engineering and malware analysis
+- Network / host / application reconnaissance
+- Defensive hardening and detection engineering
+
+Agents can be **spawned in parallel** for tasks using the built-in multi-agent orchestration.

learning/agents/agents.json

@@ -0,0 +1,7 @@
+[
+  {"id":"exploit","name":"ExploitAgent","role":"exploitation","scope":"cybersecurity-only","capabilities":["exploit analysis","payload crafting","cve triage"]},
+  {"id":"analysis","name":"AnalysisAgent","role":"analysis","scope":"cybersecurity-only","capabilities":["reverse engineering","malware analysis","traffic analysis"]},
+  {"id":"recon","name":"ReconAgent","role":"surface-mapping","scope":"cybersecurity-only","capabilities":["osint","dns","subdomain enumeration"]},
+  {"id":"vuln","name":"VulnAgent","role":"vulnerability-research","scope":"cybersecurity-only","capabilities":["vuln scanning","fuzzing","reporting"]},
+  {"id":"defense","name":"DefenseAgent","role":"defense","scope":"cybersecurity-only","capabilities":["hardening","detection","incident response"]}
+]

learning/config.json

@@ -0,0 +1,41 @@
+{
+  "model_name": "Agron",
+  "architecture": "AgronMultiAgent",
+  "model_type": "agent_system",
+  "learn_from_scratch": true,
+  "recursive_self_improvement": true,
+  "agents": [
+    {
+      "name": "ExploitAgent",
+      "role": "exploitation"
+    },
+    {
+      "name": "AnalysisAgent",
+      "role": "analysis"
+    },
+    {
+      "name": "ReconAgent",
+      "role": "surface-mapping"
+    },
+    {
+      "name": "VulnAgent",
+      "role": "vulnerability-research"
+    },
+    {
+      "name": "DefenseAgent",
+      "role": "defense"
+    }
+  ],
+  "bootstrap": {
+    "empty_memory": true,
+    "seed_rules_only": true,
+    "validation_required": true
+  },
+  "license": "MIT",
+  "domain": "cybersecurity-only",
+  "tasks": [
+    "exploitation",
+    "analysis",
+    "parallel-spawn"
+  ]
+}

learning/learning/learning_curve.json

@@ -0,0 +1,13 @@
+{
+  "model": "Agron",
+  "run": "wikitext-103-subset-50k",
+  "steps": [
+    {"step": 50, "loss": 7.461},
+    {"step": 100, "loss": 7.509},
+    {"step": 150, "loss": 7.461},
+    {"step": 200, "loss": 7.831},
+    {"step": 250, "loss": 8.117},
+    {"step": 300, "loss": 8.594}
+  ],
+  "notes": "Larger run: 8 layers, hidden 512, 1000 steps with 50k lines; early training artifacts retained"
+}

learning/learning/self_improvement.json

@@ -0,0 +1,11 @@
+{
+  "mechanisms": [
+    "self_evolve (tests + validation + optimization hints)",
+    "recon_learning extraction after completed missions",
+    "knowledge ingestion via knowledge_add/learn_url/learn_book",
+    "self_add_tool for new capabilities",
+    "install for required security binaries",
+    "git_sync auto-push schedule"
+  ],
+  "status": "enabled"
+}

learning/memory/learning_state.json

@@ -0,0 +1,9 @@
+{
+  "generation": 2,
+  "episodes": 2,
+  "knowledge_entries": 16384,
+  "improvements": [
+    "trained ~?M-param LLaMA from scratch on 50k wikitext-103 lines"
+  ],
+  "status": "trained-bigger"
+}