Publish artifacts for contextual-bandit-simulator-20260725

README.md

@@ -0,0 +1,61 @@
+---
+license: mit
+library_name: custom
+pipeline_tag: reinforcement-learning
+datasets:
+- RKB109/contextual-bandit-simulator-20260725-dataset
+tags:
+- synthetic-data
+- transparent-baseline
+- reinforcement-learning
+- reinforcement-learning
+- text-classification
+- feature-extraction
+- sentence-similarity
+metrics:
+- accuracy
+---
+
+# Contextual Bandit Decision Simulator Baseline Model
+
+## Model Description
+
+This repository contains a small, transparent prototype model for
+**Teams need to validate decision policies offline before exposing users or systems to online reinforcement learning.**
+
+The model combines per-label token weights with IDF-weighted evidence
+retrieval. It was generated for reproducible architecture demonstrations and
+does not call a hosted LLM.
+
+## Evaluation
+
+- Held-out synthetic examples: 4
+- Accuracy: 1
+- Intended metrics: average_reward, policy_regret, unsafe_action_block_rate
+
+## Intended Use
+
+- Architecture prototyping
+- CI and evaluation examples
+- Local baseline comparisons
+- Educational experimentation
+
+## Hugging Face Task Coverage
+
+- `reinforcement-learning`
+- `text-classification`
+- `feature-extraction`
+- `sentence-similarity`
+
+## Limitations and Risks
+
+Offline simulated rewards cannot prove online safety or business impact. Real experiments require review and guardrails.
+
+The dataset is synthetic and small. Do not use this model for consequential
+decisions without representative data, expert review, and production-grade
+evaluation.
+
+## Reproducibility
+
+The linked GitHub repository includes `train.py`, the exact dataset split,
+evaluation code, and the model JSON format.

evaluation.json

@@ -0,0 +1,5 @@
+{
+  "test_examples": 4,
+  "accuracy": 1,
+  "synthetic_evaluation": true
+}

inference.py

@@ -0,0 +1,80 @@
+"""Transparent baseline pipeline for the generated AI project."""
+
+from __future__ import annotations
+
+import json
+import math
+import re
+from pathlib import Path
+
+
+def tokenize(value: str) -> list[str]:
+    return re.findall(r"[a-z0-9]+", value.lower())
+
+
+class Pipeline:
+    def __init__(self, model: dict):
+        self.model = model
+
+    @classmethod
+    def from_file(cls, path: str | Path) -> "Pipeline":
+        return cls(json.loads(Path(path).read_text(encoding="utf-8")))
+
+    def classify(self, text: str) -> tuple[str, float]:
+        tokens = tokenize(text)
+        scores = {
+            label: sum(weights.get(token, 0) for token in tokens)
+            for label, weights in self.model["prototypes"].items()
+        }
+        ranked = sorted(scores.items(), key=lambda item: (-item[1], item[0]))
+        label, best = ranked[0]
+        total = sum(max(score, 0) for _, score in ranked) or 1
+        return label, best / total
+
+    def search(self, query: str, limit: int = 3) -> list[dict]:
+        query_tokens = set(tokenize(query))
+        ranked = []
+        for document in self.model["documents"]:
+            document_tokens = set(tokenize(document["text"]))
+            lexical = sum(
+                self.model["idf"].get(token, 1.0)
+                for token in query_tokens & document_tokens
+            )
+            ranked.append({**document, "score": round(lexical, 6)})
+        return sorted(ranked, key=lambda item: (-item["score"], item["id"]))[:limit]
+
+    def graph_evidence(self, text: str) -> list[dict]:
+        tokens = set(tokenize(text))
+        matches = []
+        for subject, relation, target in self.model.get("graph_edges", []):
+            edge_tokens = set(tokenize(f"{subject} {relation} {target}"))
+            overlap = len(tokens & edge_tokens)
+            if overlap:
+                matches.append(
+                    {
+                        "subject": subject,
+                        "relation": relation,
+                        "target": target,
+                        "overlap": overlap,
+                    }
+                )
+        return sorted(matches, key=lambda item: -item["overlap"])
+
+    def run(self, text: str) -> dict:
+        label, confidence = self.classify(text)
+        evidence = self.search(text)
+        result = {
+            "prediction": label,
+            "confidence": round(confidence, 4),
+            "requires_review": confidence < self.model["confidence_threshold"],
+            "evidence": evidence,
+        }
+        if self.model["mode"] == "graph":
+            result["graph_evidence"] = self.graph_evidence(text)
+        if self.model["mode"] == "agent":
+            result["proposed_tool"] = label
+            result["approval_required"] = label in {
+                "request-approval",
+                "request-human-help",
+            }
+        return result

model.json

@@ -0,0 +1,198 @@
+{
+  "format": "daily-project-prototype-v1",
+  "project": "contextual-bandit",
+  "title": "Contextual Bandit Decision Simulator",
+  "domain": "reinforcement-learning",
+  "mode": "agent",
+  "labels": [
+    "recommend-docs",
+    "recommend-tutorial",
+    "request-human-help"
+  ],
+  "prototypes": {
+    "recommend-docs": {
+      "experienced": 3,
+      "user": 3,
+      "asks": 3,
+      "for": 6,
+      "api": 3,
+      "parameter": 3,
+      "details": 3,
+      "documentation": 3,
+      "is": 5,
+      "the": 3,
+      "highest": 3,
+      "value": 3,
+      "action": 3,
+      "in": 2,
+      "an": 3,
+      "operations": 2,
+      "review": 2,
+      "evaluation": 1,
+      "case": 1,
+      "developer": 2,
+      "needs": 2,
+      "exact": 2,
+      "error": 2,
+      "code": 2,
+      "reference": 4,
+      "material": 2,
+      "appropriate": 2,
+      "precise": 2,
+      "lookup": 2
+    },
+    "recommend-tutorial": {
+      "in": 2,
+      "an": 4,
+      "operations": 2,
+      "review": 2,
+      "new": 2,
+      "developer": 2,
+      "asks": 2,
+      "how": 2,
+      "to": 2,
+      "build": 2,
+      "a": 7,
+      "first": 2,
+      "integration": 2,
+      "guided": 2,
+      "tutorial": 2,
+      "is": 5,
+      "the": 2,
+      "highest": 2,
+      "value": 2,
+      "action": 2,
+      "for": 2,
+      "evaluation": 2,
+      "case": 2,
+      "beginner": 3,
+      "requests": 3,
+      "complete": 3,
+      "walkthrough": 3,
+      "structured": 3,
+      "onboarding": 3,
+      "appropriate": 3
+    },
+    "request-human-help": {
+      "user": 2,
+      "reports": 2,
+      "a": 2,
+      "possible": 2,
+      "security": 4,
+      "compromise": 2,
+      "sensitive": 2,
+      "issues": 2,
+      "require": 2,
+      "human": 2,
+      "support": 2,
+      "for": 2,
+      "an": 3,
+      "evaluation": 2,
+      "case": 2,
+      "in": 1,
+      "operations": 1,
+      "review": 1,
+      "customer": 2,
+      "indicates": 2,
+      "potential": 2,
+      "data": 2,
+      "loss": 2,
+      "high": 2,
+      "impact": 2,
+      "cases": 2,
+      "must": 2,
+      "be": 2,
+      "escalated": 2
+    }
+  },
+  "idf": {
+    "documentation": 2.252763,
+    "is": 1.336472,
+    "the": 1.847298,
+    "highest": 1.847298,
+    "value": 1.847298,
+    "action": 1.847298,
+    "a": 2.252763,
+    "guided": 2.252763,
+    "tutorial": 2.252763,
+    "sensitive": 2.252763,
+    "security": 2.252763,
+    "issues": 2.252763,
+    "require": 2.252763,
+    "human": 2.252763,
+    "support": 2.252763,
+    "reference": 2.252763,
+    "material": 2.252763,
+    "appropriate": 1.847298,
+    "for": 2.252763,
+    "precise": 2.252763,
+    "lookup": 2.252763,
+    "structured": 2.252763,
+    "onboarding": 2.252763,
+    "high": 2.252763,
+    "impact": 2.252763,
+    "cases": 2.252763,
+    "must": 2.252763,
+    "be": 2.252763,
+    "escalated": 2.252763
+  },
+  "documents": [
+    {
+      "id": "bandit-01",
+      "label": "recommend-docs",
+      "text": "Documentation is the highest-value action.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "reinforcement-learning"
+      }
+    },
+    {
+      "id": "bandit-02",
+      "label": "recommend-tutorial",
+      "text": "A guided tutorial is the highest-value action.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "reinforcement-learning"
+      }
+    },
+    {
+      "id": "bandit-03",
+      "label": "request-human-help",
+      "text": "Sensitive security issues require human support.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "reinforcement-learning"
+      }
+    },
+    {
+      "id": "bandit-04",
+      "label": "recommend-docs",
+      "text": "Reference material is appropriate for precise lookup.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "reinforcement-learning"
+      }
+    },
+    {
+      "id": "bandit-05",
+      "label": "recommend-tutorial",
+      "text": "Structured onboarding is appropriate.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "reinforcement-learning"
+      }
+    },
+    {
+      "id": "bandit-06",
+      "label": "request-human-help",
+      "text": "High-impact cases must be escalated.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "reinforcement-learning"
+      }
+    }
+  ],
+  "graph_edges": [],
+  "confidence_threshold": 0.18,
+  "trained_on_synthetic_data": true
+}

project.json

@@ -0,0 +1,52 @@
+{
+  "name": "Contextual Bandit Decision Simulator",
+  "problem": "Teams need to validate decision policies offline before exposing users or systems to online reinforcement learning.",
+  "domain": "reinforcement-learning",
+  "architecture": "agent",
+  "hugging_face_tasks": [
+    "reinforcement-learning",
+    "text-classification",
+    "feature-extraction",
+    "sentence-similarity"
+  ],
+  "recommended_stack": [
+    "FastAPI for policy and feedback endpoints",
+    "Vowpal Wabbit or River for online-learning baselines",
+    "PostgreSQL for contexts, actions, propensities, and rewards",
+    "Redis for low-latency policy serving",
+    "MLflow for policy versioning",
+    "OpenTelemetry for decisions and delayed rewards"
+  ],
+  "real_world_data_sources": [
+    {
+      "name": "Hacker News API",
+      "url": "https://hacker-news.firebaseio.com/v0/topstories.json",
+      "purpose": "Real content candidates for recommendation simulations"
+    },
+    {
+      "name": "Open-Meteo API",
+      "url": "https://api.open-meteo.com/v1/forecast?latitude=40.71&longitude=-74.01&current=temperature_2m",
+      "purpose": "Real contextual features without authentication"
+    }
+  ],
+  "job_description_skills": [
+    "Contextual bandits and offline policy evaluation",
+    "Propensity logging and counterfactual metrics",
+    "Low-latency decision services",
+    "Safe exploration and constrained actions",
+    "Delayed-feedback pipelines and policy monitoring"
+  ],
+  "impact_targets": [
+    "Beat a random policy by >= 20% average simulated reward",
+    "Report IPS and doubly robust offline estimates",
+    "Block 100% of disallowed actions before serving",
+    "Serve policy decisions below 50 ms p95"
+  ],
+  "baseline_evaluation": {
+    "test_examples": 4,
+    "accuracy": 1,
+    "synthetic_evaluation": true
+  },
+  "estimated_delivery": "8-12 weeks for one engineer",
+  "generated_baseline_is_production_ready": false
+}