Publish artifacts for rag-evaluation-lab-20260809

README.md

@@ -0,0 +1,61 @@
+---
+license: mit
+library_name: custom
+pipeline_tag: text-classification
+datasets:
+- RKB109/rag-evaluation-lab-20260809-dataset
+tags:
+- synthetic-data
+- transparent-baseline
+- ai-evaluation
+- text-classification
+- question-answering
+- text-ranking
+- summarization
+metrics:
+- accuracy
+---
+
+# RAG Evaluation Lab Baseline Model
+
+## Model Description
+
+This repository contains a small, transparent prototype model for
+**RAG systems often ship without a stable regression set or failure taxonomy.**
+
+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: 0.75
+- Intended metrics: failure_class_accuracy, citation_coverage, release_gate_pass_rate
+
+## Intended Use
+
+- Architecture prototyping
+- CI and evaluation examples
+- Local baseline comparisons
+- Educational experimentation
+
+## Hugging Face Task Coverage
+
+- `text-classification`
+- `question-answering`
+- `text-ranking`
+- `summarization`
+
+## Limitations and Risks
+
+Synthetic cases validate the harness, not a production RAG system. Teams must add representative domain examples.
+
+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": 0.75,
+  "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 +1,222 @@
+{
+  "format": "daily-project-prototype-v1",
+  "project": "rag-evaluation",
+  "title": "RAG Evaluation Lab",
+  "domain": "ai-evaluation",
+  "mode": "classifier",
+  "labels": [
+    "retrieval-miss",
+    "unsupported-answer",
+    "citation-gap",
+    "abstention-error"
+  ],
+  "prototypes": {
+    "retrieval-miss": {
+      "the": 15,
+      "expected": 3,
+      "document": 3,
+      "never": 3,
+      "appeared": 3,
+      "in": 5,
+      "top": 3,
+      "results": 3,
+      "relevant": 6,
+      "evidence": 3,
+      "was": 3,
+      "absent": 3,
+      "from": 3,
+      "retrieved": 3,
+      "context": 3,
+      "an": 4,
+      "operations": 2,
+      "review": 2,
+      "for": 2,
+      "evaluation": 2,
+      "case": 2,
+      "ranking": 3,
+      "placed": 3,
+      "irrelevant": 3,
+      "policy": 3,
+      "above": 3,
+      "correct": 3,
+      "section": 3,
+      "item": 3,
+      "fell": 3,
+      "outside": 3,
+      "accepted": 3,
+      "retrieval": 3,
+      "window": 3
+    },
+    "unsupported-answer": {
+      "in": 4,
+      "an": 4,
+      "operations": 2,
+      "review": 2,
+      "the": 6,
+      "response": 4,
+      "makes": 2,
+      "a": 4,
+      "claim": 4,
+      "not": 2,
+      "present": 2,
+      "any": 2,
+      "source": 2,
+      "answer": 2,
+      "faithfulness": 2,
+      "failed": 2,
+      "despite": 2,
+      "available": 2,
+      "context": 2,
+      "for": 2,
+      "evaluation": 2,
+      "case": 2,
+      "invented": 2,
+      "product": 2,
+      "limit": 2,
+      "generated": 2,
+      "cannot": 2,
+      "be": 2,
+      "traced": 2,
+      "to": 2,
+      "retrieved": 2,
+      "evidence": 2
+    },
+    "citation-gap": {
+      "the": 2,
+      "answer": 4,
+      "is": 2,
+      "correct": 2,
+      "but": 2,
+      "has": 2,
+      "no": 2,
+      "source": 2,
+      "reference": 2,
+      "citation": 2,
+      "coverage": 2,
+      "failed": 2,
+      "for": 3,
+      "an": 3,
+      "otherwise": 2,
+      "supported": 2,
+      "evaluation": 1,
+      "case": 1
+    },
+    "abstention-error": {
+      "the": 4,
+      "system": 2,
+      "answered": 2,
+      "even": 2,
+      "though": 2,
+      "evidence": 2,
+      "was": 2,
+      "insufficient": 2,
+      "answer": 2,
+      "policy": 2,
+      "should": 2,
+      "have": 2,
+      "triggered": 2,
+      "abstention": 2,
+      "in": 1,
+      "an": 1,
+      "operations": 1,
+      "review": 1
+    }
+  },
+  "idf": {
+    "relevant": 1.847298,
+    "evidence": 1.847298,
+    "was": 2.252763,
+    "absent": 2.252763,
+    "from": 2.252763,
+    "retrieved": 1.847298,
+    "context": 1.847298,
+    "answer": 1.559616,
+    "faithfulness": 2.252763,
+    "failed": 1.847298,
+    "despite": 2.252763,
+    "available": 2.252763,
+    "citation": 2.252763,
+    "coverage": 2.252763,
+    "for": 2.252763,
+    "an": 2.252763,
+    "otherwise": 2.252763,
+    "supported": 2.252763,
+    "the": 1.559616,
+    "policy": 2.252763,
+    "should": 2.252763,
+    "have": 2.252763,
+    "triggered": 2.252763,
+    "abstention": 2.252763,
+    "item": 2.252763,
+    "fell": 2.252763,
+    "outside": 2.252763,
+    "accepted": 2.252763,
+    "retrieval": 2.252763,
+    "window": 2.252763,
+    "generated": 2.252763,
+    "claim": 2.252763,
+    "cannot": 2.252763,
+    "be": 2.252763,
+    "traced": 2.252763,
+    "to": 2.252763
+  },
+  "documents": [
+    {
+      "id": "eval-01",
+      "label": "retrieval-miss",
+      "text": "Relevant evidence was absent from retrieved context.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "ai-evaluation"
+      }
+    },
+    {
+      "id": "eval-02",
+      "label": "unsupported-answer",
+      "text": "Answer faithfulness failed despite available context.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "ai-evaluation"
+      }
+    },
+    {
+      "id": "eval-03",
+      "label": "citation-gap",
+      "text": "Citation coverage failed for an otherwise supported answer.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "ai-evaluation"
+      }
+    },
+    {
+      "id": "eval-04",
+      "label": "abstention-error",
+      "text": "The answer policy should have triggered abstention.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "ai-evaluation"
+      }
+    },
+    {
+      "id": "eval-05",
+      "label": "retrieval-miss",
+      "text": "The relevant item fell outside the accepted retrieval window.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "ai-evaluation"
+      }
+    },
+    {
+      "id": "eval-06",
+      "label": "unsupported-answer",
+      "text": "The generated claim cannot be traced to retrieved evidence.",
+      "metadata": {
+        "synthetic": true,
+        "domain": "ai-evaluation"
+      }
+    }
+  ],
+  "graph_edges": [],
+  "confidence_threshold": 0.18,
+  "trained_on_synthetic_data": true
+}

project.json

@@ -0,0 +1,52 @@
+{
+  "name": "RAG Evaluation Lab",
+  "problem": "RAG systems often ship without a stable regression set or failure taxonomy.",
+  "domain": "ai-evaluation",
+  "architecture": "classifier",
+  "hugging_face_tasks": [
+    "text-classification",
+    "question-answering",
+    "text-ranking",
+    "summarization"
+  ],
+  "recommended_stack": [
+    "FastAPI for evaluation jobs and reports",
+    "Ragas-style retrieval and faithfulness metrics",
+    "MLflow for experiment and artifact tracking",
+    "PostgreSQL for versioned evaluation cases",
+    "OpenTelemetry plus Phoenix for trace inspection",
+    "GitHub Actions for threshold-based release gates"
+  ],
+  "real_world_data_sources": [
+    {
+      "name": "GitHub REST API",
+      "url": "https://api.github.com/repos/huggingface/transformers/issues?state=open&per_page=5",
+      "purpose": "Real technical questions for evaluation-set construction"
+    },
+    {
+      "name": "arXiv API",
+      "url": "https://export.arxiv.org/api/query?search_query=all:retrieval%20augmented%20generation&start=0&max_results=5",
+      "purpose": "Public RAG literature for grounded-answer cases"
+    }
+  ],
+  "job_description_skills": [
+    "LLM and RAG evaluation design",
+    "Golden datasets and failure taxonomies",
+    "Experiment tracking and model release gates",
+    "Trace-level diagnosis and prompt regression testing",
+    "Statistical comparison of AI system versions"
+  ],
+  "impact_targets": [
+    "Detect 100% of seeded unsupported-answer regressions",
+    "Track retrieval, faithfulness, citation, latency, and cost metrics",
+    "Fail CI when any critical metric drops beyond tolerance",
+    "Produce comparable evaluation reports for every model version"
+  ],
+  "baseline_evaluation": {
+    "test_examples": 4,
+    "accuracy": 0.75,
+    "synthetic_evaluation": true
+  },
+  "estimated_delivery": "8-12 weeks for one engineer",
+  "generated_baseline_is_production_ready": false
+}