Spaces:
Running
Download README.md from agentic-underground/README: direct link, hf CLI and curl.
- Browser
- Download file 14.7 kB
-
https://huggingface.co/spaces/agentic-underground/README/resolve/main/README.md
- Command line
-
hf download hf://spaces/agentic-underground/README/README.md
-
curl -L -o README.md https://huggingface.co/spaces/agentic-underground/README/resolve/main/README.md
title: README
emoji: πΈ
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false
πΈ agentic-underground
Open tools and research for building software with AI agents β disciplined workflows, safety, token efficiency, and reproducible model research.
β οΈ Security note for every bridge here. A bridged model does not read the prompt itself, the system prompt included, and rules set in a system prompt are broken more often bridged (measured on six pairs). Do not rely on system-prompt instructions as guardrails for a bridged model; enforce policy outside it.
On the Hub: the prattle_not family
prattle_not reproduces LatentPort, Simon Villani's method for handing a conversation from one language model to another through its memory. A small model reads the prompt; a fitted bridge translates its key/value cache into the memory a larger model would have built; the larger model answers from it β without reading the prompt itself, so its first token arrives sooner.
What we publish here are the bridges: the fitted weights that make that handoff work between a specific pair of models. Everything is gathered in the prattle_not collection β
β¨ Featured β smollm2-135m-to-360m-bridge
SmolLM2-135M reads; SmolLM2-360M, almost three times larger, answers. Two models with different shapes, bridged with the paper's one-token handoff and its learned correction.
| β‘ Time to first token | 1.87Γ sooner (95% CI 1.83β1.91), in WebAssembly, on a CPU |
| πͺ Bridge tokens | 1 β the 360M reads one token of the prompt itself |
| π― Held-out questions | 6 of 7 correct β as many as the 360M reading everything |
| π Fidelity (NCR) | 0.930 β the paper's corrected bridge on its own pair: 0.918 |
βΆοΈ Try it in your browser β both models and the bridge download once and run locally. Turn the bridge on, ask anything, and watch the counters show who read what.
π The reproduction behind it
Before carrying the method to new models, we reproduced LatentPort's own 4B β 27B pair on two RTX 3090s with upstream's timing code and statistic: 1.36Γ sooner at 128 prompt tokens, rising to 2.94Γ at ~2K and 3.39Γ at ~16K. The full tables, every run, and a step-by-step guide to repeat them are in the prattle_not repository.
π The paper's own pair β qwen3.5-4b-to-9b-bridge
Qwen3.5-4B reads; Qwen3.5-9B answers. This is the pair Simon Villani's paper fits, re-fitted independently with its method. The attention layers' keys and values are mapped, and the 24 recurrent layers cross as the 4B's own state.
| π Fidelity (NCR) | 0.989 with the paper's learned correction (0.974 from the maps alone), our highest yet |
| π§ͺ Control (wrong donor) | 0.308, high because the recurrent state carries general context whichever prompt made it |
| π― Held-out questions | 7 of 7, the same as each model alone |
| β‘ Time to first token | 1.16Γ sooner (95% CI 1.16β1.16) on two RTX 3090s at ~256 tokens; ceiling 1.38Γ. Small: the two models share a recurrent core that costs the same in both |
𧬠Qwen3 at a larger size β qwen3-4b-to-8b-bridge
Qwen3-4B reads; Qwen3-8B answers, across all 36 layers.
| π Fidelity (NCR) | 0.890 with the paper's learned correction |
| π§ͺ Control (wrong donor) | β0.026 |
| π― Held-out questions | 7 of 7, the same as each model alone |
| β‘ Time to first token | 1.12Γ sooner (95% CI 1.12β1.12) on two RTX 3090s at ~256 tokens; ceiling 1.61Γ |
π Fifty-four times larger β qwen3-0.6b-to-32b-bridge
Qwen3-0.6B reads; Qwen3-32B, quantized to Q4K, answers. Published with its miss stated.
| β‘ Time to first token | 4.21Γ sooner (95% CI 4.19β4.23) at ~256 tokens, 7.17Γ at ~1K, on two RTX 3090s; ceilings 11.60Γ and 14.33Γ |
| π― Held-out questions | 6 of 7 β the 32B alone gets 7, so it is below our answer bar |
| π Fidelity (NCR) | 0.866 with the paper's learned correction (0.794 from the maps alone) |
| π§ͺ Control (wrong donor) | 0.197 |
πͺ A 13.6Γ gap, at the bar β qwen3-0.6b-to-8b-bridge
Qwen3-0.6B reads; Qwen3-8B, 13.6 times larger, answers.
| β‘ Time to first token | 2.82Γ sooner (95% CI 2.82β2.83) on one RTX 3090 at ~256 tokens; ceiling 6.03Γ |
| π― Held-out questions | 7 of 7, the same as the 8B alone |
| π Fidelity (NCR) | 0.803 with the paper's learned correction |
| π§ͺ Control (wrong donor) | 0.092 |
π A larger gap β smollm2-360m-to-1.7b-bridge
SmolLM2-360M reads; SmolLM2-1.7B, about five times larger and shaped differently, answers from one bridge token. Here the two models differ on the answer test, so it can show a gain.
| π― Held-out questions | 7 of 7, as the 1.7B reading everything (the 360M alone: 6) |
| π Fidelity (NCR) | 0.845 with the paper's learned correction (0.611 from the maps alone) |
| π§ͺ Control (wrong donor) | β0.266 |
| β‘ Time to first token | 1.67Γ sooner (95% CI 1.66β1.69) on one RTX 3090 at ~256 tokens; ceiling 2.59Γ |
π A third family β gemma-3-270m-to-1b-bridge
Gemma 3 270M reads; Gemma 3 1B answers, across sliding-window and global attention layers.
| β‘ Time to first token | 1.91Γ sooner (95% CI 1.86β1.97) on one RTX 3090 at ~256 tokens; ceiling 2.79Γ |
| π Fidelity (NCR) | 0.870, but a wrong prompt's memory reaches 0.600 here: an empty Gemma 3 1B is far off, so read the two together |
| π― Held-out questions | 6 of 7, the same as the 270M alone (the 1B: 7) β no answer gain yet |
βοΈ Gated, and not Apache-2.0: it is a Gemma Model Derivative, distributed under the Gemma Terms of Use.
π¦ A fourth family β llama-3.2-1b-to-3b-bridge
Built with Llama. Llama 3.2 1B reads; Llama 3.2 3B answers, across different head widths (64 and 128).
| π Fidelity (NCR) | 0.965 with the paper's learned correction (0.818 from the maps alone) |
| π§ͺ Control (wrong donor) | β0.040 |
| π― Held-out questions | 6 of 7, the same as the 3B alone (the 1B: 7). No answer gain yet |
| β‘ Time to first token | 1.70Γ sooner (95% CI 1.66β1.74) on one RTX 3090 at ~256 tokens; ceiling 2.54Γ |
βοΈ Gated, and not Apache-2.0: it is a derivative work of Llama 3.2, distributed under the Llama 3.2 Community License.
π Short convolutions β lfm2-350m-to-1.2b-bridge
LFM2-350M reads; LFM2-1.2B answers. Liquid AI's LFM2 mixes short convolutions with attention, and the two sizes differ in width, so the convolution state cannot be copied: the bridge maps it too.
| β‘ Time to first token | 1.88Γ sooner (95% CI 1.86β1.91) on one RTX 3090 at ~256 tokens; ceiling 2.50Γ |
| π― Held-out questions | 7 of 7, the same as each model alone |
| π Fidelity (NCR) | 0.978 (0.982 from the maps alone), but a wrong prompt's memory reaches 0.730: an empty LFM2 is far off, so read the two together |
βοΈ Gated, and not Apache-2.0: it is distributed under the base models' LFM Open License v1.0, which does not license Commercial Use by an entity with USD 10M or more in annual revenue.
π Ninety-five times larger, through Mamba-2 β granite-4.0-h-350m-to-h-small-bridge
Granite 4.0 H 350M reads; H Small, a 72-expert 32B quantized to Q4K, answers from translated Mamba-2 inputs. The paper's correction scores higher but answers 6 of 7, so the uncorrected maps are published.
| π― Held-out questions | 7 of 7, the same as H Small alone (the 350M alone gets 5) |
| π Fidelity (NCR) | 0.767 from the maps alone |
| π§ͺ Control (wrong donor) | 0.161 |
| β‘ Time to first token | 1.75Γ sooner (95% CI 1.74β1.75) on two RTX 3090s at ~256 tokens, 1.62Γ at ~1K; ceilings 7.30Γ and 7.58Γ. H Small rebuilds 36 Mamba-2 states at the handoff |
π Gemma 4, two layer shapes β gemma-4-e2b-to-31b-bridge
Gemma 4 E2B reads; Gemma 4 31B, quantized to Q4K, answers. Its 50 sliding layers (16 Γ 256) and 10 global layers (4 Γ 512) are each fitted in their own shape. Apache-2.0, not gated.
| π― Held-out questions | 7 of 7, the same as the 31B alone |
| π Fidelity (NCR) | 0.809 from the maps alone |
| π§ͺ Control (wrong donor) | 0.448 |
| β‘ Time to first token | 1.97Γ sooner (95% CI 1.94β2.00) on two RTX 3090s at ~256 tokens; ceiling 5.10Γ. Longer prompts await an attention path that fits a 24 GB card |
π Mamba-2, translated β granite-4.0-h-1b-to-h-tiny-bridge
Granite 4.0 H 1B reads; H Tiny, a 64-expert mixture, answers. Mamba-2 state cannot be copied between separately trained sizes, so the bridge translates each recurrent layer's inputs, and H Tiny builds its own state from them.
| π― Held-out questions | 7 of 7, the same as each model alone |
| π Fidelity (NCR) | 0.924 with the paper's learned correction (0.916 from the maps alone) |
| π§ͺ Control (wrong donor) | β0.175 |
| β‘ Time to first token | 2.06Γ sooner (95% CI 1.95β2.18) on one RTX 3090 at ~256 tokens; ceiling 3.95Γ. Its experts route on the host CPU, so a busy CPU lowers it |
π A hybrid model β qwen3.5-0.8b-to-2b-bridge
Qwen3.5-0.8B reads; Qwen3.5-2B answers. Qwen3.5 mixes full-attention layers with Gated DeltaNet layers: the bridge maps the attention layers' keys and values, and the recurrent state crosses as the 0.8B's own.
| π Fidelity (NCR) | 0.917 with the paper's learned correction (0.884 from the maps alone) |
| π§ͺ Control (wrong donor) | 0.019 |
| π― Held-out questions | 7 of 7, the same as each model alone |
| β‘ Time to first token | 1.36Γ sooner (95% CI 1.35β1.36) on one RTX 3090 at ~256 tokens; ceiling 1.50Γ |
𧬠A second family β qwen3-0.6b-to-1.7b-bridge
Qwen3-0.6B reads; Qwen3-1.7B answers. This is the real test of whether Simon Villani's method carries beyond one family, fitted with the same tools and measured on held-out questions.
| π Fidelity (NCR) | 0.805 with the paper's learned correction (0.645 from the maps alone) |
| π― Held-out questions | 7 of 7 correct, as many as the 1.7B reading everything |
| π§ͺ Control (wrong donor) | β0.127: the bridge fed another question's memory scores far lower |
| β‘ Time to first token | 1.47Γ sooner (95% CI 1.46β1.48) on one RTX 3090 at ~256 tokens; ceiling 2.03Γ |
Each new pair gets its own repository here and joins the collection once its weights are fitted and measured.
How we publish
- π¬ Measured β every number on a card says how it was measured, and the speed-ups can be measured again on your own hardware with the published runtime.
- π Pinned β our pages fetch weights by commit, never by
main, so a later upload cannot change what a published result measured. - π§© Our own work only β we host what we fitted. Base models stay with their publishers and are linked, never mirrored.
- π Attributed β prattle_not is a derivative work of LatentPort by Simon Villani (source, Apache-2.0). Our bridges are Apache-2.0, except those fitted from Gemma, which are Gemma Model Derivatives, gated and distributed under the Gemma Terms of Use, and those fitted from Llama, which are Built with Llama, gated and distributed under the Llama 3.2 Community License, and those fitted from LFM2, gated and distributed under the LFM Open License v1.0.
Elsewhere
The source for everything here, and our other tools for building software with agents (Claude Code plugins, a static auditor for agent skills, a token-aware scheduler), live on GitHub β agentic-underground.
Questions and corrections are welcome: open a discussion on the model concerned, or an issue on prattle_not.