CoolFace
Apppublic

Patron82/ai-agents-course-local-workflow

sourceHugging Faceupdated 15d agoView on Hugging Face
0likes
App README

AI Agents Course — documented local workflow

Course exercise prepared with Codex assistance for Patron82. This is a transparent, source-assisted learning project, not an independent personal examination or a blind GAIA benchmark.

The public Space hosts source code and a receipt from the local model run. The Space does not run the MLX model. The verified run was performed on Apple Silicon with Qwen2.5-Coder-7B-Instruct-4bit, MLX, LangGraph, deterministic tools and local BGE-small embeddings. No paid inference service is required.

Method and limitations

  • —One frozen workflow processes all 20 official course questions.
  • —The source catalog and factual notes were researched with Codex. Source discovery is not automated and this prior research limits benchmark generalization.
  • —The agent performs executable counting, table comparison and food-organ classification, model-based extraction from documented facts, and query/retrieve/extract over two live original sources.
  • —General answer-format conversions expand a city abbreviation and extract documented family names. No task-ID-specific expected-answer substitutions occur.
  • —Five original task attachments were unavailable from the course API (HTTP 404). Their answers remain null locally and are submitted as empty strings. All 20 tasks remain in the evaluation denominator.
  • —run_results.json contains actual outputs from one full local run. It is a cache of generated outputs, not the agent implementation or a ground-truth answer key. No official review answer list is imported by the workflow.
  • —Local comparison found 15 source-supported answers. Only the official scoring endpoint determines the official score. Historical date snapshots were not verified for every source; see qualifications in evidence.json.

Reproduce on Apple Silicon

Use Python 3.11+, install requirements-local.txt, and obtain mlx-community/Qwen2.5-Coder-7B-Instruct-4bit in a local model directory. Populate a FastEmbed cache for BAAI/bge-small-en-v1.5 first. The embedding reader is deliberately local-only.

Set AGENT_MODEL_PATH to the model directory and EMBEDDING_CACHE to the FastEmbed cache. PDF_PYTHON optionally selects a Python interpreter with pypdf; it defaults to the active interpreter. Run:

sh
python integrated_agent.py --run-id my-reproduction

Results and tool/model traces are written under runs/; newly downloaded source documents under live_sources/. A run ID cannot overwrite an existing run. Live pages and dependency updates can change later results. Source and package hashes for the recorded run are in run_results.json.

Files

integrated_agent.py: orchestration and generic formatting. structured_source_workflow.py: query/retrieve/extract and literal evidence validation. extensions.py: deterministic tools and bounded public source retrieval. semantic_passages.py: local semantic passage ranking. JSON files contain official questions or attributed source facts. Original source documents are not mirrored by this public Space.

Attribution

Course and task API: https://huggingface.co/learn/agents-course/en/unit4/hands-on

Public source links and limitations are recorded in the source-fact JSON files. Third-party materials retain their original rights. This source publication documents the course exercise; it does not assert formal accreditation, an employment qualification, or unaided work.