PUMA Vault

Ollama — Local LLM Inference Engine

Se lee en 5 min

Ollama — Local LLM Inference Engine

Purpose in PUMA: Ollama is the inference backend that runs Llama 3.2 8B and Mistral 7B locally without GPU or API costs, enabling reproducible, deterministic experiments (seed=42, temperature=0).

Key Properties

PropertyValue
Installcurl -fsSL https://ollama.ai/install.sh | sh
APILocal REST API on localhost:11434
Modelsollama pull llama3.2:8b / ollama pull mistral:7b
Reproducibilityseed=42, temperature=0 via API params
HardwareRuns on CPU with 16GB RAM (quantized 4-bit)
RAM usage~5GB for 8B model (Q4_K_M quantization)

Basic API Call (Python)

import requests
import json
 
def call_ollama(prompt: str, model: str = "llama3.2:8b") -> str:
    """
    Call Ollama local API with deterministic settings.
    Seed and temperature=0 ensure reproducibility.
    """
    response = requests.post(
        "http://localhost:11434/api/generate",
        json={
            "model": model,
            "prompt": prompt,
            "stream": False,
            "options": {
                "seed": 42,
                "temperature": 0,
                "num_predict": 512
            }
        }
    )
    return response.json()["response"]

Model Versions Used in PUMA

ModelTagRAMSpeed (CPU)Use
Llama 3.2llama3.2:8b~5GB~15-40s/queryPrimary
Mistral 7Bmistral:7b~4.5GB~12-35s/queryComparison
Phi-3.5 Miniphi3.5:3.8b~2.5GB~6-15s/queryFallback (latency)

🔗 Related Notes

30 - Permanent/31 Concepts/PN-LLM-Agents | SP-Architecture 60 - Resources/61 Prompts/PT-PUMA-Triage-ZeroShot

id: LN-Tool-ClaudeCode title: “Claude Code — Agentic Coding CLI” type: literature-note subtype: tool tags: [tool, claude-code, coding-agent, cli, anthropic] url: “https://docs.claude.ai/claude-code” puma_role: “Primary AI coding assistant for implementation” puma_phase: “F1, F2, F3” created: 2026-03-01

Claude Code — Agentic Coding CLI

Purpose in PUMA: Claude Code is used for implementing the PUMA benchmark modules (triage agent, estimation agent, metric calculator), generating test cases from BDD specs, and refactoring code to meet reproducibility standards.

Key Capabilities

  • Reads and edits files directly in the project
  • Runs tests and interprets failures autonomously
  • Understands full codebase context
  • Follows SDD specs when provided as input

Effective Prompt Patterns for PUMA

Pattern 1 — Spec-to-Code Generation

# In terminal with Claude Code active:
claude "Read the spec at docs/specs/SP-Triage-Agent-v1.md and implement 
the TriageAgent class in src/agents/triage_agent.py. 
Requirements:
- Use Ollama API at localhost:11434
- Accept model name and strategy as parameters
- Return structured JSON with prediction and reasoning
- Add type hints and docstrings
- Include unit test skeleton in tests/test_triage_agent.py"

Pattern 2 — Debug with Context

claude "The Wilcoxon test in src/analysis/stats.py is producing p-values 
above 1.0. Here is the error: [paste error]. 
Look at the input data format and fix the issue. 
Explain what was wrong before making changes."

Pattern 3 — Reproducibility Audit

claude "Audit the entire src/ directory for reproducibility issues:
1. Find any random calls without seed=42
2. Find any hardcoded paths that should be configurable
3. Find any missing version pins in requirements.txt
Report findings as a checklist. Do not make changes yet."

🔗 Related Notes

LN-Tool-OpenCode | 60 - Resources/61 Prompts/61.3 Dev-Tools/PT-ClaudeCode-Agent-Triage

id: LN-Tool-OpenCode title: “OpenCode — Open-Source AI Coding Agent” type: literature-note subtype: tool tags: [tool, opencode, coding-agent, open-source] github: “https://github.com/opencode-ai/opencode” puma_role: “Open-source alternative to Claude Code for code generation” puma_phase: “F2, F3” created: 2026-03-01

OpenCode — Open-Source AI Coding Agent

Purpose in PUMA: OpenCode provides an open-source, locally-runnable alternative to Claude Code — important for the project’s commitment to open-source tooling and reproducibility without API costs.

Configuration for PUMA

# opencode.yaml
model: "ollama/llama3.2:8b"  # or any Ollama model
context:
  include:
    - "src/**/*.py"
    - "docs/specs/*.md"
    - "tests/**/*.py"
tools:
  - read_file
  - write_file
  - run_command
  - search_files

Best Prompt Patterns

# Structured task with context
[CONTEXT]: Working on PUMA benchmark — a reproducible LLM evaluation 
framework for ICT project management.

[TASK]: Implement the function calculate_f1_macro() in 
src/metrics/classification.py

[SPEC]: 
- Input: y_true (list[str]), y_pred (list[str]), classes (list[str])
- Output: dict with keys: f1_macro, f1_per_class, precision, recall
- Use sklearn.metrics internally
- Add comprehensive docstring with example

[CONSTRAINTS]:
- Python 3.11
- No external deps beyond scikit-learn
- Type hints required

🔗 Related Notes

LN-Tool-ClaudeCode | 60 - Resources/61 Prompts/61.3 Dev-Tools/PT-OpenCode-Refactor

id: LN-Tool-BrowserOS title: “Browser OS — AI Web Browsing Agent” type: literature-note subtype: tool tags: [tool, browser-os, web-agent, automation] puma_role: “Automated web research and dataset discovery” puma_phase: “F0, F1” created: 2026-03-01

Browser OS — AI Web Browsing Agent

Purpose in PUMA: Browser OS automates web-based research tasks: navigating academic databases, downloading papers, checking dataset availability, and verifying DOIs — freeing researcher time for higher-value analysis.

Use Cases in PUMA

TaskPrompt Pattern
Check paper availability”Navigate to [DOI URL] and confirm the paper is open-access or provide alternatives”
Dataset verification”Go to zenodo.org/records/5901893 and confirm: file format, size, licence, last update”
Citation count check”Search Semantic Scholar for [paper title] and report citation count and recent citing papers”
Tool documentation”Navigate to ollama.ai/docs and find the API parameter for controlling random seed”

Example PUMA Prompt

ROLE: You are a research assistant with web access.
CONTEXT: I am working on PUMA Project about LLM agents for project management.
OBJECTIVE: Find the 3 most recent papers (2024-2026) that benchmark 
           local LLMs on software engineering classification tasks.
INSTRUCTIONS:
1. Search arXiv.org with query: "local LLM benchmark software engineering 2024"
2. Filter for papers with available code repositories
3. For each paper: extract title, authors, date, arXiv ID, GitHub link
FORMAT: Markdown table with 5 columns: Title | Authors | Date | arXiv | GitHub

🔗 Related Notes

60 - Resources/61 Prompts/61.3 Dev-Tools/PT-BrowserOS-Web-Agent

Vista Gráfica

Retroenlaces