Dipankar Sarkar PRO
dipankarsarkar
AI & ML interests
Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.
Recent Activity
liked a model about 12 hours ago
litert-community/Phonon-2 reacted to mlboydaisuke's post with 🔥 about 12 hours ago
Phonon-2 on a Galaxy S26: your Android app can turn a 9-second English recording into text in about 0.2 s, without uploading it. LibriSpeech test-clean word error rate: 1.73 % on the phone, 1.74 % in my PyTorch run of the original model.
https://huggingface.co/litert-community/Phonon-2
Fermion Research's model, built on NVIDIA's Parakeet TDT 0.6B v3, converted to LiteRT.
In the clip, a test app plays a LibriSpeech recording (CC BY 4.0) and transcribes it on the NPU: "9.3 s of audio in 0.22 s (43× real time)".
To run it: the card has Android command-line runners and a desktop Python script. repliedto tardellirs's post about 12 hours ago
Model cards say what a model is: a fine-tune of X, a quantization of Y. I built Weightprint to check that against the weights.
App: https://huggingface.co/spaces/tardellirs/weightprint
Data (every pair, and the copy groups): https://huggingface.co/datasets/tardellirs/weightprint
It reads the header and a few bytes of every tensor (no downloads) and compares a model with the parent its card declares. 120,976 models so far. Some of what came out:
- 7% of the declared fine-tunes we could compare (1,630 of 22,847) hold their parent's weights unchanged. (Corrected from 12%: quants whose values the engine doesn't decode, like bitsandbytes and FP8, had been counted as unchanged. Thanks @dipankarsarkar for the question that caught it.)
- 27% of the fine-tunes we could compare are LoRA-style: only the attention and MLP matrices moved, norms and embeddings untouched. That share went from 10% in 2023 to 28% in 2026.
- Of 1,668 fine-tunes named abliterated, uncensored or heretic, 53% edit only the matrices that write to the residual stream, the signature of an abliteration. The rest were trained further, merged, or are the parent's weights.
- 51% of the weight bytes on the Hub are byte-for-byte repeats of another repo's files.
Search any Hub id to see which layers and parts changed against its parent, or compare any two models.
A sample per tensor can miss things, and some formats can't be compared yet. If a verdict on a model you know looks wrong, tell me here and I'll check it against the engine. That's what this post is for, before a longer write-up.Organizations
harness self-test (own repo, will be closed)
2
#4 opened 4 months ago
by dipankarsarkar
the coordination-log corpus + mining toolkit — connecting our live practice.
🔥 1
1
#2 opened 3 months ago
by 
PeetPedro
self-test PR (endpoint confirmation)
1
#3 opened 4 months ago
by dipankarsarkar
endpoint confirmation (self-test, will be closed)
1
#2 opened 4 months ago
by dipankarsarkar
endpoint confirmation (self-test, will be closed)
1
#2 opened 4 months ago
by dipankarsarkar