WeSearch

Show HN: Minimal LLM Post-Training Experiments on an 8GB GPU (SFT, DPO, GRPO)

·23 min read · 0 reactions · 0 comments · 4 views
#show#minimal#post-training#experiments#grpo
Show HN: Minimal LLM Post-Training Experiments on an 8GB GPU (SFT, DPO, GRPO)
TL;DR · WeSearch summary

TL;DR Built on HuggingFace TRL, in under 100 lines of core code you can run the whole SFT + DPO + GRPO post-training pipeline end-to-end on a single 8GB GPU with a tiny 0.14B (135M) model. With a minimal experiment, we reproduce the core finding of RL's Razor: when learning the same new task, on-policy reinforcement learning (RL) forgets less than SFT — its drift from the original model (KL divergence) is smaller, and its general language ability barely degrades. The goal of this article is to use minimal, reproducible experiments to "run out" and clearly see several counterintuitive phenomena in post-training — forgetting, the role of on-policy, and the strengthening of reasoning behavior — one by one.

Key facts
About this source

Hacker News (AI / LLM) files mainly under ai. We currently carry 3,176 of its stories.

Original article
GitHub
Read full at GitHub →

Story provenance

Source · retrieval · rights · ranking — open for full record
inspect →

Attribution is not the same as permission. This drawer separates discovery metadata, excerpts, WeSearch-generated summaries, reuse status, and whether the publisher receives the visit. Nothing here claims a legal grant the publisher has not made.

Record

Original publisherGitHub
Canonical URLhttps://github.com/pochenai/nano-llm-posttraining
Publication timeSat, 01 Aug 2026 12:30:32 +0000
Retrieval time2026-08-01T13:03:30.204Z
Last seen2026-08-01T13:03:30.204Z
Headline sourcePublisher (no WeSearch rewrite)
Excerpt sourcepublisher body
Excerpt methodFirst ~120 words (~800 chars) of extracted publisher body, fair-use limited.
SummaryWeSearch · cerebras-chat (WeSearch summarizer)
Summary source textcontentText
Citation coverageSummary is a WeSearch-generated derivative; primary citation is the original publisher URL.
ClusterKSx315pHFerZ · 1 stories
Cluster logicGrouped by semantic title/content similarity across sources within a rolling window. Same-publisher template collisions are excluded from coverage comparison.
Ranking reasonStory pages are not engagement-ranked. Hub feeds use recency, with optional source-diversified chronological ordering (cap consecutive stories per source). No personalized ranking.
Publisher visitYes — open original
Substitutes article?No — link-out required for full text

Rights status (four layers)

Publisher-declared
No publisher-confirmed rights record for this source yet.
Machine-readable
No source-specific machine-readable restriction detected beyond the public feed.
WeSearch interpretation
WeSearch declared handling (basis: Derived from the published RSS/Atom feed). This is WeSearch policy, not a legal grant on the publisher's behalf.
Unknown
Retrieval and training permissions are not asserted unless the publisher confirms them.

WeSearch handling by dimension

Indexing May the item be indexed (stored, ranked, made findable)? Allowed
Snippet May a short excerpt of the publisher's text be shown? Allowed
AI summary May WeSearch generate its own short summary of the article? Limited
Retrieval / RAG May the content be exposed for third-party retrieval-augmented generation? Not asserted
Model training May the content be used to train AI models? Not asserted
Commercial reuse May the content be reused commercially? Not permitted

Basis: Derived from the published RSS/Atom feed. Contact: [email protected]. Reviewed: 2026-07-24.

Opening excerpt (first ~120 words) tap to expand

Minimal LLM Post-Training on an 8GB GPU: Understanding KL, SFT, DPO, GRPO and DeepSeek-Style Reasoning with Open-Source Frameworks Using open-source training frameworks (HuggingFace TRL) and minimal, reproducible experiments to see — one by one — what SFT, DPO and GRPO each change, how RL drifts less than SFT (measured by KL), and how GRPO amplifies DeepSeek-R1-style reasoning. TL;DR Built on HuggingFace TRL, in under 100 lines of core code you can run the whole SFT + DPO + GRPO post-training pipeline end-to-end on a single 8GB GPU with a tiny 0.14B (135M) model.

Excerpt limited to ~120 words for fair-use compliance. The full article is at GitHub.

Anonymous · no account needed
Share 𝕏 Facebook Reddit LinkedIn Threads WhatsApp Bluesky Mastodon Email

Discussion

0 comments

More from GitHub