Spaces:
Running
Running
publish from LatentPortPort 7f79ac0
Browse files
README.md
CHANGED
|
@@ -14,8 +14,8 @@ pinned: false
|
|
| 14 |
**Open tools and research for building software with AI agents** β
|
| 15 |
disciplined workflows, safety, token efficiency, and reproducible model research.
|
| 16 |
|
| 17 |
-
<a href="https://huggingface.co/collections/agentic-underground/
|
| 18 |
-
<a href="https://agentic-underground.github.io/
|
| 19 |
<a href="https://github.com/agentic-underground"><img alt="Source on GitHub" src="https://img.shields.io/badge/source-GitHub-24292f?logo=github"></a>
|
| 20 |
<a href="https://huggingface.co/papers/2609.25053"><img alt="Builds on LatentPort by Simon Villani (arXiv 2609.25053)" src="https://img.shields.io/badge/builds%20on-LatentPort%20by%20Simon%20Villani-b31b1b"></a>
|
| 21 |
|
|
@@ -27,9 +27,9 @@ disciplined workflows, safety, token efficiency, and reproducible model research
|
|
| 27 |
> prompt included, and rules set in a system prompt are broken more often bridged (measured on six pairs).
|
| 28 |
> Do not rely on system-prompt instructions as guardrails for a bridged model; enforce policy outside it.
|
| 29 |
|
| 30 |
-
## On the Hub: the
|
| 31 |
|
| 32 |
-
**
|
| 33 |
handing a conversation from one language model to another *through its memory*. A small model reads
|
| 34 |
the prompt; a fitted **bridge** translates its key/value cache into the memory a larger model would
|
| 35 |
have built; the larger model answers from it β **without reading the prompt itself**, so its first
|
|
@@ -37,7 +37,7 @@ token arrives sooner.
|
|
| 37 |
|
| 38 |
What we publish here are the **bridges**: the fitted weights that make that handoff work between a
|
| 39 |
specific pair of models. Everything is gathered in the
|
| 40 |
-
**[
|
| 41 |
|
| 42 |
### β¨ Featured β [`smollm2-135m-to-360m-bridge`](https://huggingface.co/agentic-underground/smollm2-135m-to-360m-bridge)
|
| 43 |
|
|
@@ -51,7 +51,7 @@ shapes, bridged with the paper's **one-token handoff** and its learned correctio
|
|
| 51 |
| π― **Held-out questions** | **6 of 7** correct β as many as the 360M reading everything |
|
| 52 |
| π **Fidelity (NCR)** | **0.930** β the paper's corrected bridge on its own pair: 0.918 |
|
| 53 |
|
| 54 |
-
βΆοΈ **[Try it in your browser](https://agentic-underground.github.io/
|
| 55 |
bridge download once and run locally. Turn the bridge on, ask anything, and watch the counters show
|
| 56 |
who read what.
|
| 57 |
|
|
@@ -60,7 +60,7 @@ who read what.
|
|
| 60 |
Before carrying the method to new models, we reproduced LatentPort's own **4B β 27B** pair on two
|
| 61 |
RTX 3090s with upstream's timing code and statistic: **1.36Γ** sooner at 128 prompt tokens, rising to
|
| 62 |
**2.94Γ** at ~2K and **3.39Γ** at ~16K. The full tables, every run, and a step-by-step guide to
|
| 63 |
-
repeat them are in the [
|
| 64 |
|
| 65 |
### π The paper's own pair β [`qwen3.5-4b-to-9b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-4b-to-9b-bridge)
|
| 66 |
|
|
@@ -239,7 +239,7 @@ and measured.
|
|
| 239 |
what a published result measured.
|
| 240 |
- π§© **Our own work only** β we host what we fitted. Base models stay with their publishers and are
|
| 241 |
linked, never mirrored.
|
| 242 |
-
- π **Attributed** β
|
| 243 |
Simon Villani ([source](https://github.com/LatentPort/latentport), Apache-2.0). Our bridges are
|
| 244 |
Apache-2.0, except those fitted from Gemma, which are Gemma Model Derivatives, gated and distributed
|
| 245 |
under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms), and those fitted from Llama, which
|
|
@@ -253,4 +253,4 @@ Code plugins, a static auditor for agent skills, a token-aware scheduler), live
|
|
| 253 |
**[GitHub β agentic-underground](https://github.com/agentic-underground)**.
|
| 254 |
|
| 255 |
Questions and corrections are welcome: open a discussion on the model concerned, or an
|
| 256 |
-
[issue on
|
|
|
|
| 14 |
**Open tools and research for building software with AI agents** β
|
| 15 |
disciplined workflows, safety, token efficiency, and reproducible model research.
|
| 16 |
|
| 17 |
+
<a href="https://huggingface.co/collections/agentic-underground/prattle-not-6af7ac5e95e0fc80de563d61"><img alt="Collection: prattle_not" src="https://img.shields.io/badge/collection-prattle__not-6366f1"></a>
|
| 18 |
+
<a href="https://agentic-underground.github.io/prattle_not/"><img alt="Live demo in your browser" src="https://img.shields.io/badge/demo-runs%20in%20your%20browser-a855f7"></a>
|
| 19 |
<a href="https://github.com/agentic-underground"><img alt="Source on GitHub" src="https://img.shields.io/badge/source-GitHub-24292f?logo=github"></a>
|
| 20 |
<a href="https://huggingface.co/papers/2609.25053"><img alt="Builds on LatentPort by Simon Villani (arXiv 2609.25053)" src="https://img.shields.io/badge/builds%20on-LatentPort%20by%20Simon%20Villani-b31b1b"></a>
|
| 21 |
|
|
|
|
| 27 |
> prompt included, and rules set in a system prompt are broken more often bridged (measured on six pairs).
|
| 28 |
> Do not rely on system-prompt instructions as guardrails for a bridged model; enforce policy outside it.
|
| 29 |
|
| 30 |
+
## On the Hub: the prattle_not family
|
| 31 |
|
| 32 |
+
**prattle_not** reproduces [**LatentPort**](https://huggingface.co/papers/2609.25053), Simon Villani's method for
|
| 33 |
handing a conversation from one language model to another *through its memory*. A small model reads
|
| 34 |
the prompt; a fitted **bridge** translates its key/value cache into the memory a larger model would
|
| 35 |
have built; the larger model answers from it β **without reading the prompt itself**, so its first
|
|
|
|
| 37 |
|
| 38 |
What we publish here are the **bridges**: the fitted weights that make that handoff work between a
|
| 39 |
specific pair of models. Everything is gathered in the
|
| 40 |
+
**[prattle_not collection β](https://huggingface.co/collections/agentic-underground/prattle-not-6af7ac5e95e0fc80de563d61)**
|
| 41 |
|
| 42 |
### β¨ Featured β [`smollm2-135m-to-360m-bridge`](https://huggingface.co/agentic-underground/smollm2-135m-to-360m-bridge)
|
| 43 |
|
|
|
|
| 51 |
| π― **Held-out questions** | **6 of 7** correct β as many as the 360M reading everything |
|
| 52 |
| π **Fidelity (NCR)** | **0.930** β the paper's corrected bridge on its own pair: 0.918 |
|
| 53 |
|
| 54 |
+
βΆοΈ **[Try it in your browser](https://agentic-underground.github.io/prattle_not/)** β both models and the
|
| 55 |
bridge download once and run locally. Turn the bridge on, ask anything, and watch the counters show
|
| 56 |
who read what.
|
| 57 |
|
|
|
|
| 60 |
Before carrying the method to new models, we reproduced LatentPort's own **4B β 27B** pair on two
|
| 61 |
RTX 3090s with upstream's timing code and statistic: **1.36Γ** sooner at 128 prompt tokens, rising to
|
| 62 |
**2.94Γ** at ~2K and **3.39Γ** at ~16K. The full tables, every run, and a step-by-step guide to
|
| 63 |
+
repeat them are in the [prattle_not repository](https://github.com/agentic-underground/prattle_not).
|
| 64 |
|
| 65 |
### π The paper's own pair β [`qwen3.5-4b-to-9b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-4b-to-9b-bridge)
|
| 66 |
|
|
|
|
| 239 |
what a published result measured.
|
| 240 |
- π§© **Our own work only** β we host what we fitted. Base models stay with their publishers and are
|
| 241 |
linked, never mirrored.
|
| 242 |
+
- π **Attributed** β prattle_not is a derivative work of [LatentPort](https://latentport.com/) by
|
| 243 |
Simon Villani ([source](https://github.com/LatentPort/latentport), Apache-2.0). Our bridges are
|
| 244 |
Apache-2.0, except those fitted from Gemma, which are Gemma Model Derivatives, gated and distributed
|
| 245 |
under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms), and those fitted from Llama, which
|
|
|
|
| 253 |
**[GitHub β agentic-underground](https://github.com/agentic-underground)**.
|
| 254 |
|
| 255 |
Questions and corrections are welcome: open a discussion on the model concerned, or an
|
| 256 |
+
[issue on prattle_not](https://github.com/agentic-underground/prattle_not/issues).
|