Nirav-Madhani/Meditation

Meditation: Self-Supervised Introspection as a Training Phase for Language Models

A research project exploring a structured introspection phase inserted between supervised fine-tuning and task-specific reinforcement learning.

Overview

During meditation, a model freely explores mathematical concepts, constructs its own problems, and develops observations — rewarded by a composite signal blending programmatic verification with LLM-judged novelty.

Current Status

Pipeline

SFT → Meditation RL → Task RL → Post-training (block removal)

Training Stack

• —Student model: LiquidAI/LFM2-8B-A1B (MoE, 1.5B active params)
• —Training: QLoRA (4-bit NF4), GRPO
• —Seed generation: Qwen3.5-4B via SGLang (local, GPU)
• —Judge: Qwen3.5-1.7B via llama.cpp (CPU, OpenAI-compatible API)
• —Hardware: Single NVIDIA RTX 4070 (8GB VRAM)

Evaluation Benchmarks

• —GSM8K (1,319 test problems) — primary math benchmark
• —MATH-500 (500 competition-level problems, per-subject breakdown) — harder reasoning
• —ARC-Challenge — reasoning generalization outside mathematics

All benchmarks report accuracy with and without the meditation block at inference time.

Repository Structure

configs/          Training configuration files
data/             Seed meditations (1,427 filtered) and topic curriculum (188 topics)
scripts/          Pipeline scripts (seed gen, filtering, SFT, RL, evaluation)
evals/            Evaluation outputs (populated after training)
checkpoints/      Saved model checkpoints (not tracked in git)

Paper

The working paper (meditation-paper.html) is served directly via this Space.