caveman_tokenstransfer.2.0 — a fork that adds an input-side peer (100% local LLMLingua-2) — courtesy ping

[vfalbor]

Hey @JuliusBrussee and team — first, thanks for caveman. The reframing from "be concise" to a dialect was a real insight and the numbers are clean.

After watching the discourse split into "input compression vs output compression," I built and just released caveman_tokenstransfer.2.0 — a fork that keeps caveman exactly as it is and adds two peer skills on the input side:

• /tokenstransfer — LLMLingua-2 input compression, 100% local by default (pip install llmlingua torch tiktoken), no API key needed. Same fully-local philosophy caveman has.
• /tokenstranslation — multilingual prompt → English (cheap-tokens language) for non-English UX.
• /caveman-fullstack — runs /caveman output dialect + /tokenstransfer input compression in one shot.

It's MIT, it's a real fork (linked to this repo), it ships your installer rebranded (all 30+ agents work). Your output skill is unchanged — credit goes to you.

why a fork instead of a PR

The new pieces add a Python dependency (llmlingua, torch) and a hosted-fallback service. Both feel like things you'd want to keep out of upstream caveman's "zero deps, pure prompts" surface. So I shipped it next door instead of trying to merge it in.

benchmark on your own 10 prompts

Suite	caveman only	tokenstransfer only (input)	fullstack
your 10 prompts (mostly short-Q / long-A)	-65%	-2%	-65%
RAG (long context, short answer)	-62%	-27%	-64%
agent loop (system + tools + history)	-63%	-42%	-69%

For your workload (codegen short-Q long-A) caveman alone is enough. For long-context apps the input side is where the bill is. The two compose cleanly.

Independent input-side head-to-head vs caveman-compress skill (your rule-based version) on your own benchmark: -52.9% LLMLingua-2 vs -8.6% rules. Reproduction repo: vfalbor/llm-language-token-tax · vs-caveman.

what I want from you

1. Permission gut-check: I tried to be respectful with the fork (kept your name in install banner, point installers back at upstream where they should, link to your repo in README hero). If anything reads wrong, tell me and I'll fix it.
2. Sanity-check on the framing: I'm pitching this as "you took the output side; we took the input side; stack both." If you think that frame misrepresents caveman in any way, I want to know before I do any louder promo.
3. (Optional) Upstream collaboration: if you'd ever want a caveman-transfer-lite or similar that wraps llmlingua as a soft dep inside upstream, happy to PR.

Not asking for stars — caveman is already at 64k and crushing. Just doing the polite thing of telling you next door.

Why use many token when few token do trick. 🪨🌳