beanie00 / self-distillation-analysis

What is self-distillation-analysis?

This Python codebase lets researchers reproduce and dissect self-distillation experiments on large language models, zeroing in on why it tanks reasoning performance in math tasks. Built atop SDPO for preference optimization, it handles data prep from datasets like DAPO-Math-17k, model evaluation with hints, SFT dataset creation, and training runs via GRPO or SDPO variants. Users get Docker setups, eval scripts, and Hugging Face checkpoints for Qwen and DeepSeek models—ideal for probing epistemic verbalization loss without rebuilding from scratch. Unlike flaky codebases not working in Cursor or GitHub Copilot's codebase indexing hiccups, this one prioritizes clean repro on large math problems.

Why is it gaining traction?

In a sea of LLM training repos, it stands out by pinpointing a subtle failure: confident teacher outputs suppress uncertainty signals, hurting generalization on broad task coverage. Devs dig the full pipeline—eval baselines, rich feedback loops, multiturn generalization tests—over vague workspace vs codebase debates in Copilot setups. With W&B logs and arXiv linkage, it's a go-to for analysis on GitHub codebase RAG or xAI-style large codebase GitHub experiments, even if stars are modest at 25.

Who should use this?

LLM researchers dissecting RLHF pitfalls, math reasoning trainers scaling DAPO-style datasets, or academics verifying self-distillation claims before fine-tuning. Perfect for teams hitting OOD drops in code/math hybrids, or anyone benchmarking against baselines like GRPO on GPU clusters.

Verdict

Grab it if you're deep in LLM reasoning repros—solid Dockerized flow and HF models make it practical despite 1.0% credibility from low stars and academic focus. Skip for production; it's research-grade, not battle-tested at scale. (198 words)