Skills DirectorySkills Directory
SkillsLearnSecurityCategoriesDocsCommunityBlog
Sign InSubmit Skill
Skills Directory

Security-tested agent skills for Claude, coding agents, and AI workflows.

Directory

  • Browse Skills
  • All Skills A–Z
  • Claude Skills
  • Claude Code Skills
  • Agent Skills
  • Categories
  • Submit a Skill

Learn

  • Learn Hub
  • Install Claude Skills
  • Write SKILL.md
  • Skills vs MCP
  • Directories Compared

Security

  • Security
  • Methodology
  • Secure Claude Skills
  • Security Badges

Company

  • About
  • Community
  • Blog
  • API Docs
  • Advertise

2026 Skills Directory. All rights reserved.

Back to skills

Model Reflection Tuning

ASecurity

Use when pipeline reflection-tuning (auto-crítica 2 estágios) com validação reproduzível via HashHop. Gera datasets CoT→crítica→correção, treina modelos abertos, detecta benchmarks não reproduzíveis (ex.: Reflection 70B). Triggers on \"model-reflection-tuning\", \"model reflection tuning\", \"tuning\".

2 stars
0 votes
0 copies
0 views
Added 9/19/2026
ai-agentsbashapiperformance

Works with

api

Security Analysis

A100/100

Scanned 9/19/2026

Install to Claude Code

$npx -y skills add majinmagros/magros.ai-skills --skill model-reflection-tuning --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Model Reflection Tuning?

Add the live security badge to your README — it updates automatically with every re-scan.

Security grade badge for Model Reflection Tuning
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/majinmagros-model-reflection-tuning/badge)](https://www.skillsdirectory.com/skills/majinmagros-model-reflection-tuning)

More formats (shields.io, HTML) on the badges page.

Download Zip
Files
SKILL.md
---
name: model-reflection-tuning
description: "Use when pipeline reflection-tuning (auto-crítica 2 estágios) com validação reproduzível via HashHop. Gera datasets CoT→crítica→correção, treina modelos abertos, detecta benchmarks não reproduzíveis (ex.: Reflection 70B). Triggers on \"model-reflection-tuning\", \"model reflection tuning\", \"tuning\"."
metadata:
  origin: ECC
  module: workflow-quality
  tools: Read, Write, Edit, Bash, Grep, Glob, Task
---

# Model Reflection-Tuning Skill

Pipeline completo para aplicar **reflection-tuning** (fine-tuning com auto-crítica em 2 estágios) em modelos abertos, com validação reproduzível via **HashHop** e detecção de benchmarks fraudulentos.

## Quando Ativar

- Fine-tuning de modelos abertos (Llama, Qwen, Mistral) com reflection-tuning
- Gerar datasets sintéticos: CoT → auto-crítica → correção (reward shaping)
- Validar resultados via HashHop (não confiar em benchmarks do autor)
- Detectar weights/benchmarks não reproduzíveis (caso Reflection 70B)
- Auditorar claims de performance de modelos "reflection-tuned"

## Pipeline (4 Estágios)

```
┌─────────────────┐   ┌──────────────────┐   ┌──────────────────┐   ┌──────────────────┐
│ 1. Dataset Gen  │→  │ 2. Reflection    │→  │ 3. Training      │→  │ 4. Validation    │
│    (CoT +       │   │    Tuning        │   │    (LoRA/Full)   │   │    (HashHop +    │
│     critique)   │   │    Trainer       │   │                  │   │     Reproducer)  │
└─────────────────┘   └──────────────────┘   └──────────────────┘   └──────────────────┘
```

### Estágio 1: Dataset Generation (`reflection_dataset_gen.py`)

Gera tarefas → modelo propõe+resolve → reward shaping (learnability + accuracy).

**Entradas:**
- `base_model`: modelo base (ex.: `meta-llama/Llama-3.1-8B-Instruct`)
- `task_specs`: lista de especificações de tarefas (JSON/YAML)
- `num_samples`: amostras por tarefa
- `critique_model`: modelo para auto-crítica (pode ser o próprio base_model)

**Saída:** `dataset/reflection_dataset.jsonl` com campos:
```json
{
  "task_id": "math_001",
  "prompt": "Solve: 2x + 5 = 15",
  "cot": "Step 1: Subtract 5... Step 2: Divide by 2... Answer: x=5",
  "critique": "The solution is correct but step 2 could show division explicitly",
  "corrected_cot": "Step 1: Subtract 5 from both sides... Step 2: Divide both sides by 2... Answer: x=5",
  "reward_learnability": 0.85,
  "reward_accuracy": 1.0
}
```

**Reward Shaping:**
- `learnability`: quão fácil o modelo aprende com a correção (0-1)
- `accuracy`: correção factual da resposta final (0-1)
- Combined: `0.6 * accuracy + 0.4 * learnability`

### Estágio 2: Reflection Tuning Trainer (`reflection_trainer.py`)

Treina com loss combinada: `L = L_ce + λ * L_reflection`

**Configuração:**
- `base_model`: HF model ID ou path local

Attribution

majinmagrosmajinmagros
View sourceMore from majinmagros →
SSkills DirectorySkills Directory

Ship a skill? Prove it's safe.

Free 120-pattern security scan, letter grade, and an embeddable README badge.

Submit a skill

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments (0)

No comments yet. Be the first to comment!

SSkills DirectorySkills Directory

Ship a skill? Prove it's safe.

Free 120-pattern security scan, letter grade, and an embeddable README badge.

Submit a skill

Related Skills

Caveman

Ultra-compressed communication mode. Cuts token usage ~75% by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman. Also auto-triggers when token efficiency is requested.

1023331 votes

Hyperplan

Adversarial multi-agent planning skill. Self-orchestrates 5 hostile category members (unspecified-low, unspecified-high, deep, ultrabrain, artistry) via team-mode for ruthless cross-critique debate, distills only the defensible insights, then MANDATORILY hands the distilled insight bundle to the `plan` agent for executable plan formalization. Use when planning needs maximum rigor and surfacing of weak assumptions, blind spots, and over-engineering. Triggers: 'hyperplan', 'hpp', '/hyperplan', ...

686011 votes

Mcp Code Execution

Routes multi-tool workflows through MCP servers for large datasets and pipelines. Use when Bash tool overhead is limiting throughput on data-heavy tasks.

3331 votes

catchup

Recovers prior coding-agent session context by running `catchup <agent> --since-compact`, which extracts a clean summary of a previous Codex, Claude Code, Antigravity, OpenCode, or Pi Agent session. Use when the user says "catch up", "what did the last session do", "get me up to speed", "I switched agents", or asks to recover/summarize a previous session before continuing. Do NOT use for the current conversation, git history, or any non-agent log.

611 votes

math-skill

A comprehensive mathematical reasoning skill for AI assistants — handles arithmetic to research-level problems with rigorous step-by-step reasoning, systematic verification, and transparent uncertainty handling

381 votes
View all in ai-agents →