Tonic
Β·
AI & ML interests
π€Making robots to help people learn things quicker π©π»βππ
Recent Activity
reacted to RDTvlokip's post with π about 22 hours ago I finally changed the architecture of my 15M French LLM. It worked. Then I almost fooled myself about how much and catching that was the real win.
After proving last time that architecture is a threshold, not a lever, I got stubborn: could I change how the model learns? Four honest attempts, Lion, a sharper AdamW Ξ²2, multi-token prediction, LayerScale. Four failures. The bottleneck wasn't the learning rule either.
So I changed the shape of the computation instead: loop the same transformer blocks 4Γ, deeper reasoning, zero added parameters. It beat the baseline on perplexity, the first thing in the whole project to move that number. Then I added my own twist: let each token decide how deep to think, halting on its own entropy.
My first evaluation was spectacular. Coherence up 65%. Hallucinated names down 62%.
It was noise.
Eight prompts, one seed. I re-ran on 50 prompts Γ 200 tokens and watched the gains shrink to "modest" and on out-of-domain prompts, recurrence actually made things worse. No universal winner. And none of it is new: it's Adaptive Computation Time (2016), the Universal Transformer (2018), and LoopViT (2026), recombined and measured honestly.
The real lesson:
A number from 8 prompts is a rumor. The eval harness that kills your own best result is worth more than the result it kills. Cite your lineage. Stay preliminary until multiple seeds say otherwise.
The three models are live. The write-up is honest about every caveat π
π https://huggingface.co/blog/RDTvlokip/teaching-a-15m-french-llm-to-think-deeper Organizations
Tonic/landshift-sft-v1-duplicate
Viewer
β’ Updated β’ 239 β’ 26
Tonic/brief-composer-sft-v1-duplicate
Viewer
β’ Updated β’ 3k β’ 80
Tonic/oceanscout-sft-v1-duplicate
Viewer
β’ Updated β’ 69 β’ 26
Tonic/nutonic-sft-init-upload-test
Viewer
β’ Updated β’ 310 β’ 158
Viewer
β’ Updated β’ 1.04k β’ 40
Preview
β’ Updated β’ 85
Tonic/android-operator-text-full
Viewer
β’ Updated β’ 9.91k β’ 8
Tonic/voxtral-dataset-20260225_222829
Viewer
β’ Updated β’ 10 β’ 7
Tonic/voxtral-dataset-20250913_175651
Viewer
β’ Updated β’ 10 β’ 5
Tonic/voxtral-dataset-20250913_174653
Viewer
β’ Updated β’ 10 β’ 7
Tonic/android-operator-episodes
Updated β’ 47
β’ 2
Viewer
β’ Updated β’ 929 β’ 9
Tonic/trackio-experiments
Viewer
β’ Updated β’ 32 β’ 333
β’ 1
Tonic/finreg_dataset_gemma3_27b_it_qat
Viewer
β’ Updated β’ 5 β’ 11
Tonic/ollama_1000_example
Viewer
β’ Updated β’ 1.05k β’ 35
Viewer
β’ Updated β’ 263 β’ 18
β’ 1
Tonic/Health-Bench-Eval-OSS-2025-07
Viewer
β’ Updated β’ 9.67k β’ 113
β’ 4
Tonic/twitter-block-lists
Viewer
β’ Updated β’ 6.69k β’ 6
Tonic/scaleway_r1_dark_thoughts_casestudies_processed_fuzzy_think_splits
Viewer
β’ Updated β’ 904 β’ 4
β’ 1
Tonic/runpod_qwen32_dark_thoughts_casestudies_processed_fuzzy_think_splits
Viewer
β’ Updated β’ 78.6k β’ 18
Tonic/scaleway_r1_dark_thoughts_casestudies
Viewer
β’ Updated β’ 1.3M β’ 35
β’ 2
Viewer
β’ Updated β’ 56.9k β’ 23
β’ 3
Tonic/dark_thoughts_stakeholders_deduplicated_shard
Viewer
β’ Updated β’ 1.78k β’ 5
β’ 3
Tonic/dark_thoughts_stakeholders_test
Viewer
β’ Updated β’ 689k β’ 15
Tonic/Climate-Guard-Toxic-Agent
Viewer
β’ Updated β’ 83.8k β’ 105
β’ 2
Viewer
β’ Updated β’ 488 β’ 408
β’ 7
Tonic/combined-fr-caselaw-21-01-2025
Updated β’ 130
Tonic/combined-fr-caselaw
Viewer
β’ Updated β’ 511k β’ 28
Updated β’ 2
Viewer
β’ Updated β’ 15.5k β’ 23
β’ 2
![Joseph [open/acc] Pollack's picture](https://cdn-avatars.huggingface.co/v1/production/uploads/62a3bb1cd0d8c2c2169f0b88/eT2TS0IlQbZtz-F_zHLz9.jpeg)
