Full write-up with the architecture, the answer checks and the lessons from running it publicly:
https://huggingface.co/blog/Mati83moni/kos-a-polish-voice-assistant-that-is-allowed-to-sa
Mateusz Piesiak PRO
Mati83moni
AI & ML interests
Hybrid LLMs, MoE, Mamba, multilingual NLP, biomedical reasoning, efficient pretraining
Recent Activity
updated a Space 3 days ago
Mati83moni/kos-pl upvoted a changelog 8 days ago
External Collaborators from Repo Settings repliedto their post 8 days ago
Kos: a Polish voice assistant that's allowed to say "I'm not sure" 🐦
Kos is a free Polish voice assistant running on ZeroGPU: Silero VAD → Whisper large-v3 → PLLuM-12B (EXL3) → TTS. It's a cascade, not a native speech-to-speech model.
The interesting part is what happens between "the model wrote a sentence" and "the sentence is spoken":
• math, units, exchange rates (NBP) and weather are computed by code, not by the LLM
• facts go to Polish Wikipedia/Wikidata first, with sources you can check
• free-chat answers are buffered and rejected if they contain an unsourced number, a fake citation or a false "I have no internet"
• a rejected answer is never spoken: Kos looks it up or says it isn't sure
Voice: my own cloned voice on Cartesia. When its daily/monthly limit runs out, Kos switches automatically to Chatterbox, then Piper. It still answers, it just sounds different.
Limits: it's a 12B model, so small talk can still be wrong. Polish only. ZeroGPU quota is shared by all visitors.
Try it (allow the mic, speak Polish, headphones work best):
https://huggingface.co/spaces/Mati83moni/kos-pl
If it answers something wrong, tell me in the Space's Community tab: what you said, what Kos replied. Those cases become regression tests.Organizations
None yet
replied to their post 8 days ago
posted an update 8 days ago
Post
108
Kos: a Polish voice assistant that's allowed to say "I'm not sure" 🐦
Kos is a free Polish voice assistant running on ZeroGPU: Silero VAD → Whisper large-v3 → PLLuM-12B (EXL3) → TTS. It's a cascade, not a native speech-to-speech model.
The interesting part is what happens between "the model wrote a sentence" and "the sentence is spoken":
• math, units, exchange rates (NBP) and weather are computed by code, not by the LLM
• facts go to Polish Wikipedia/Wikidata first, with sources you can check
• free-chat answers are buffered and rejected if they contain an unsourced number, a fake citation or a false "I have no internet"
• a rejected answer is never spoken: Kos looks it up or says it isn't sure
Voice: my own cloned voice on Cartesia. When its daily/monthly limit runs out, Kos switches automatically to Chatterbox, then Piper. It still answers, it just sounds different.
Limits: it's a 12B model, so small talk can still be wrong. Polish only. ZeroGPU quota is shared by all visitors.
Try it (allow the mic, speak Polish, headphones work best):
Mati83moni/kos-pl
If it answers something wrong, tell me in the Space's Community tab: what you said, what Kos replied. Those cases become regression tests.
Kos is a free Polish voice assistant running on ZeroGPU: Silero VAD → Whisper large-v3 → PLLuM-12B (EXL3) → TTS. It's a cascade, not a native speech-to-speech model.
The interesting part is what happens between "the model wrote a sentence" and "the sentence is spoken":
• math, units, exchange rates (NBP) and weather are computed by code, not by the LLM
• facts go to Polish Wikipedia/Wikidata first, with sources you can check
• free-chat answers are buffered and rejected if they contain an unsourced number, a fake citation or a false "I have no internet"
• a rejected answer is never spoken: Kos looks it up or says it isn't sure
Voice: my own cloned voice on Cartesia. When its daily/monthly limit runs out, Kos switches automatically to Chatterbox, then Piper. It still answers, it just sounds different.
Limits: it's a 12B model, so small talk can still be wrong. Polish only. ZeroGPU quota is shared by all visitors.
Try it (allow the mic, speak Polish, headphones work best):
Mati83moni/kos-pl
If it answers something wrong, tell me in the Space's Community tab: what you said, what Kos replied. Those cases become regression tests.
- 1 reply
reacted to their post with 🔥 14 days ago
Post
5,007
Is "evaluation awareness" in a language model one direction, or eight?
We held the task fixed, toggled the eight cue factors of EvalAwareBench one at a time, and looked at the residual stream of Qwen3 (0.6B–32B), Gemma 3 and PLLuM-12B, plus a first Polish cross-lingual pilot.
What we found (descriptive, same instrument and author throughout):
- Cue directions overlap at about 1/5–1/4 of what two estimates of the same cue reach: not one shared axis.
- The standard all-cues direction is mostly an explicit-disclosure detector (cosine 0.81–0.89 with that cue).
- It transfers poorly to stimulus artificiality (held-out AUROC ≈ 0.5–0.6), though the models do represent that cue.
- No statistically detectable MMLU framing effect in the tested settings. That does not establish absence of an effect or of sandbagging.
- The Polish pilot is inconclusive: the formal verdict says so, and its own length control was not met.
The article carries its corrections (three early findings were withdrawn at scale; one traced to a bug in our hook code), an open validation problem for the add interventions, and all 23 figures with sources.
Full article: https://huggingface.co/blog/Mati83moni/is-evaluation-awareness-one-direction-or-eight
Spaces:
Mati83moni/evaluation-awareness-one-direction-or-eight
Mati83moni/evaluation-awareness-pllum-12b
We held the task fixed, toggled the eight cue factors of EvalAwareBench one at a time, and looked at the residual stream of Qwen3 (0.6B–32B), Gemma 3 and PLLuM-12B, plus a first Polish cross-lingual pilot.
What we found (descriptive, same instrument and author throughout):
- Cue directions overlap at about 1/5–1/4 of what two estimates of the same cue reach: not one shared axis.
- The standard all-cues direction is mostly an explicit-disclosure detector (cosine 0.81–0.89 with that cue).
- It transfers poorly to stimulus artificiality (held-out AUROC ≈ 0.5–0.6), though the models do represent that cue.
- No statistically detectable MMLU framing effect in the tested settings. That does not establish absence of an effect or of sandbagging.
- The Polish pilot is inconclusive: the formal verdict says so, and its own length control was not met.
The article carries its corrections (three early findings were withdrawn at scale; one traced to a bug in our hook code), an open validation problem for the add interventions, and all 23 figures with sources.
Full article: https://huggingface.co/blog/Mati83moni/is-evaluation-awareness-one-direction-or-eight
Spaces:
Mati83moni/evaluation-awareness-one-direction-or-eight
Mati83moni/evaluation-awareness-pllum-12b
- 2 replies
reacted to yanghaojin's post with 🚀 14 days ago
Post
5,612
DeepSeek-V4.1-Flash-1017, running locally on a Mac
Our 4-bit paged build is up: GreenBitAI/DeepSeek-V4.1-Flash-1017-4bit-paged
On one M3 Ultra Mac Studio with gbx:
• ~22 tok/s standard decode
• ~44 tok/s on code with DeepSeek's official draft head (1.99×)
• ~33 tok/s on chat with the draft head
Why run 1017: DeepSeek reports 77.0 on DeepSWE v1.1 (up from 74.2) and 42.3 on Terminal-Bench 4.0 (up from 31.2), with shorter outputs (https://api-docs.deepseek.com/news/news261017).
It reads text and images, and plugs straight into Claude Code and Codex; the setup is in the model card. The paged format also loads on Macs with less memory, and on an NVIDIA DGX Spark with experts paged from the SSD;
Engine: https://github.com/GreenBitAI/gbx-lm/releases
Our 4-bit paged build is up: GreenBitAI/DeepSeek-V4.1-Flash-1017-4bit-paged
On one M3 Ultra Mac Studio with gbx:
• ~22 tok/s standard decode
• ~44 tok/s on code with DeepSeek's official draft head (1.99×)
• ~33 tok/s on chat with the draft head
Why run 1017: DeepSeek reports 77.0 on DeepSWE v1.1 (up from 74.2) and 42.3 on Terminal-Bench 4.0 (up from 31.2), with shorter outputs (https://api-docs.deepseek.com/news/news261017).
It reads text and images, and plugs straight into Claude Code and Codex; the setup is in the model card. The paged format also loads on Macs with less memory, and on an NVIDIA DGX Spark with experts paged from the SSD;
Engine: https://github.com/GreenBitAI/gbx-lm/releases
reacted to DedeProGames's post with 🧠🚀🔥 14 days ago
Post
8,923
🚀 Introducing Kiyo-230M-Instruct — my compact English & Portuguese chat model!
Built on Kiyo-230M-Preview, a base model I pretrained from scratch, Kiyo now has an instruction-tuned release for everyday conversations, rewriting, and lightweight assistant experiments.
🧠 229.7M parameters
🌍 English + Portuguese
💬 Native chat template for multi-turn conversations
📏 2,048-token context window
📦 459 MB of BF16 safetensors weights
🔓 Apache-2.0 license
For instruction tuning, I used 165,735 conversations from SmolTalk2, smol-smoltalk, UltraChat, and Dolci-Instruct-SFT. Training ran for two complete epochs with assistant-only loss, using a roughly 70% English / 30% Portuguese token mix.
The updated model card includes benchmarks against SmolLM2-135M-Instruct, BananaMind, and Supra2, plus example conversations, CPU measurements, training details, and a Transformers quickstart.
The results are mixed: Kiyo is competitive with these small models across the reported evaluations, while SmolLM2 and BananaMind score higher on the instruction-following benchmarks. Portuguese was a highlight in the small conversational sanity test.
This is an experimental 230M model, with plenty of room to improve reasoning, factual accuracy, and strict instruction following. I’m excited to share it with people exploring how much we can get from small, accessible models!
🤗 Model & full README:
DedeProGames/Kiyo-230M-Instruct
💬 Try the chat demo:
DedeProGames/kiyo-230m-chat
Try it in English or Portuguese — I’d love to see your examples, feedback, and experiments!
Built on Kiyo-230M-Preview, a base model I pretrained from scratch, Kiyo now has an instruction-tuned release for everyday conversations, rewriting, and lightweight assistant experiments.
🧠 229.7M parameters
🌍 English + Portuguese
💬 Native chat template for multi-turn conversations
📏 2,048-token context window
📦 459 MB of BF16 safetensors weights
🔓 Apache-2.0 license
For instruction tuning, I used 165,735 conversations from SmolTalk2, smol-smoltalk, UltraChat, and Dolci-Instruct-SFT. Training ran for two complete epochs with assistant-only loss, using a roughly 70% English / 30% Portuguese token mix.
The updated model card includes benchmarks against SmolLM2-135M-Instruct, BananaMind, and Supra2, plus example conversations, CPU measurements, training details, and a Transformers quickstart.
The results are mixed: Kiyo is competitive with these small models across the reported evaluations, while SmolLM2 and BananaMind score higher on the instruction-following benchmarks. Portuguese was a highlight in the small conversational sanity test.
This is an experimental 230M model, with plenty of room to improve reasoning, factual accuracy, and strict instruction following. I’m excited to share it with people exploring how much we can get from small, accessible models!
🤗 Model & full README:
DedeProGames/Kiyo-230M-Instruct
💬 Try the chat demo:
DedeProGames/kiyo-230m-chat
Try it in English or Portuguese — I’d love to see your examples, feedback, and experiments!
- 9 replies
reacted to alibidaran's post with 🔥 14 days ago
Post
3,874
I have recently worked on a personal project called progressive thinking, which enables LLMs to generate chains of thought in a progressive pattern. In addition, the model learned to generate COT before choosing the appropriate tools, making it useful for complex and deeper agentic tasks. This model is well trained on QWEN3-14B using the SFT algorithm and needs GRPO to make feasible outputs.
alibidaran/qwen_progressive_thinking_16bit
OpenDataArena/Spark-234K
IFM/TxT360-3efforts
alibidaran/qwen_progressive_thinking_16bit
OpenDataArena/Spark-234K
IFM/TxT360-3efforts
reacted to danielhanchen's post with 🔥 14 days ago
Post
5,025
Run DeepSeek V4.1-Flash-1017 locally! 🐳
We made 1-bit retain ~88.2% accuracy while shrinking size from 508GB to 250GB (−52%).
Run on a 128GB+ RAM device. V4.1 Flash 1017 is the strongest open model to date
Guide: https://unsloth.ai/docs/models/deepseek-v4.1-flash
GGUF: unsloth/DeepSeek-V4.1-Flash-1017-GGUF
We made 1-bit retain ~88.2% accuracy while shrinking size from 508GB to 250GB (−52%).
Run on a 128GB+ RAM device. V4.1 Flash 1017 is the strongest open model to date
Guide: https://unsloth.ai/docs/models/deepseek-v4.1-flash
GGUF: unsloth/DeepSeek-V4.1-Flash-1017-GGUF
replied to their post 14 days ago
Thanks a lot ! Appreciate you taking the time to read it.
reacted to their post with ❤️ 15 days ago
Post
5,007
Is "evaluation awareness" in a language model one direction, or eight?
We held the task fixed, toggled the eight cue factors of EvalAwareBench one at a time, and looked at the residual stream of Qwen3 (0.6B–32B), Gemma 3 and PLLuM-12B, plus a first Polish cross-lingual pilot.
What we found (descriptive, same instrument and author throughout):
- Cue directions overlap at about 1/5–1/4 of what two estimates of the same cue reach: not one shared axis.
- The standard all-cues direction is mostly an explicit-disclosure detector (cosine 0.81–0.89 with that cue).
- It transfers poorly to stimulus artificiality (held-out AUROC ≈ 0.5–0.6), though the models do represent that cue.
- No statistically detectable MMLU framing effect in the tested settings. That does not establish absence of an effect or of sandbagging.
- The Polish pilot is inconclusive: the formal verdict says so, and its own length control was not met.
The article carries its corrections (three early findings were withdrawn at scale; one traced to a bug in our hook code), an open validation problem for the add interventions, and all 23 figures with sources.
Full article: https://huggingface.co/blog/Mati83moni/is-evaluation-awareness-one-direction-or-eight
Spaces:
Mati83moni/evaluation-awareness-one-direction-or-eight
Mati83moni/evaluation-awareness-pllum-12b
We held the task fixed, toggled the eight cue factors of EvalAwareBench one at a time, and looked at the residual stream of Qwen3 (0.6B–32B), Gemma 3 and PLLuM-12B, plus a first Polish cross-lingual pilot.
What we found (descriptive, same instrument and author throughout):
- Cue directions overlap at about 1/5–1/4 of what two estimates of the same cue reach: not one shared axis.
- The standard all-cues direction is mostly an explicit-disclosure detector (cosine 0.81–0.89 with that cue).
- It transfers poorly to stimulus artificiality (held-out AUROC ≈ 0.5–0.6), though the models do represent that cue.
- No statistically detectable MMLU framing effect in the tested settings. That does not establish absence of an effect or of sandbagging.
- The Polish pilot is inconclusive: the formal verdict says so, and its own length control was not met.
The article carries its corrections (three early findings were withdrawn at scale; one traced to a bug in our hook code), an open validation problem for the add interventions, and all 23 figures with sources.
Full article: https://huggingface.co/blog/Mati83moni/is-evaluation-awareness-one-direction-or-eight
Spaces:
Mati83moni/evaluation-awareness-one-direction-or-eight
Mati83moni/evaluation-awareness-pllum-12b
- 2 replies
posted an update 15 days ago
Post
5,007
Is "evaluation awareness" in a language model one direction, or eight?
We held the task fixed, toggled the eight cue factors of EvalAwareBench one at a time, and looked at the residual stream of Qwen3 (0.6B–32B), Gemma 3 and PLLuM-12B, plus a first Polish cross-lingual pilot.
What we found (descriptive, same instrument and author throughout):
- Cue directions overlap at about 1/5–1/4 of what two estimates of the same cue reach: not one shared axis.
- The standard all-cues direction is mostly an explicit-disclosure detector (cosine 0.81–0.89 with that cue).
- It transfers poorly to stimulus artificiality (held-out AUROC ≈ 0.5–0.6), though the models do represent that cue.
- No statistically detectable MMLU framing effect in the tested settings. That does not establish absence of an effect or of sandbagging.
- The Polish pilot is inconclusive: the formal verdict says so, and its own length control was not met.
The article carries its corrections (three early findings were withdrawn at scale; one traced to a bug in our hook code), an open validation problem for the add interventions, and all 23 figures with sources.
Full article: https://huggingface.co/blog/Mati83moni/is-evaluation-awareness-one-direction-or-eight
Spaces:
Mati83moni/evaluation-awareness-one-direction-or-eight
Mati83moni/evaluation-awareness-pllum-12b
We held the task fixed, toggled the eight cue factors of EvalAwareBench one at a time, and looked at the residual stream of Qwen3 (0.6B–32B), Gemma 3 and PLLuM-12B, plus a first Polish cross-lingual pilot.
What we found (descriptive, same instrument and author throughout):
- Cue directions overlap at about 1/5–1/4 of what two estimates of the same cue reach: not one shared axis.
- The standard all-cues direction is mostly an explicit-disclosure detector (cosine 0.81–0.89 with that cue).
- It transfers poorly to stimulus artificiality (held-out AUROC ≈ 0.5–0.6), though the models do represent that cue.
- No statistically detectable MMLU framing effect in the tested settings. That does not establish absence of an effect or of sandbagging.
- The Polish pilot is inconclusive: the formal verdict says so, and its own length control was not met.
The article carries its corrections (three early findings were withdrawn at scale; one traced to a bug in our hook code), an open validation problem for the add interventions, and all 23 figures with sources.
Full article: https://huggingface.co/blog/Mati83moni/is-evaluation-awareness-one-direction-or-eight
Spaces:
Mati83moni/evaluation-awareness-one-direction-or-eight
Mati83moni/evaluation-awareness-pllum-12b
- 2 replies
reacted to their post with 👍❤️ about 1 month ago
Post
71
🔍 Watermark Finder — find, mark, and strip hidden text watermarks. Open source, MIT.
Detects six distinct covert channels with exact character offsets, not guesses: Unicode tag payloads, variation-selector encoding, zero-width binary/markers, homoglyph substitution, bidi controls, and repetition stamps. Also verifies C2PA content credentials (Adobe, Leica, Google-signed manifests) — and keeps "integrity" (did the bytes change?) strictly separate from "trust" (is the signer who they claim?), because a forged certificate verifies perfectly too.
It scores writing register (how much prose reads like unedited assistant output) with visible uncertainty ranges and per-feature contributions — and refuses a verdict below ~15 words rather than fake precision. That score measures register, not authorship, and the README says so louder than I'm saying it here.
Two deploy modes: a hosted engine (FastAPI on a HF Space) behind a Cloudflare Worker, or a fully client-side build compiled to WebAssembly via Pyodide — the document never leaves the browser tab in that mode.
304 tests across four suites, real CI on every push, no accounts (anonymous HMAC-token workspace).
Live:
https://watermark-finder.pages.dev/
Source:
https://github.com/Mati83mon/Watermark_Finder
Detects six distinct covert channels with exact character offsets, not guesses: Unicode tag payloads, variation-selector encoding, zero-width binary/markers, homoglyph substitution, bidi controls, and repetition stamps. Also verifies C2PA content credentials (Adobe, Leica, Google-signed manifests) — and keeps "integrity" (did the bytes change?) strictly separate from "trust" (is the signer who they claim?), because a forged certificate verifies perfectly too.
It scores writing register (how much prose reads like unedited assistant output) with visible uncertainty ranges and per-feature contributions — and refuses a verdict below ~15 words rather than fake precision. That score measures register, not authorship, and the README says so louder than I'm saying it here.
Two deploy modes: a hosted engine (FastAPI on a HF Space) behind a Cloudflare Worker, or a fully client-side build compiled to WebAssembly via Pyodide — the document never leaves the browser tab in that mode.
304 tests across four suites, real CI on every push, no accounts (anonymous HMAC-token workspace).
Live:
https://watermark-finder.pages.dev/
Source:
https://github.com/Mati83mon/Watermark_Finder
posted an update about 2 months ago
Post
71
🔍 Watermark Finder — find, mark, and strip hidden text watermarks. Open source, MIT.
Detects six distinct covert channels with exact character offsets, not guesses: Unicode tag payloads, variation-selector encoding, zero-width binary/markers, homoglyph substitution, bidi controls, and repetition stamps. Also verifies C2PA content credentials (Adobe, Leica, Google-signed manifests) — and keeps "integrity" (did the bytes change?) strictly separate from "trust" (is the signer who they claim?), because a forged certificate verifies perfectly too.
It scores writing register (how much prose reads like unedited assistant output) with visible uncertainty ranges and per-feature contributions — and refuses a verdict below ~15 words rather than fake precision. That score measures register, not authorship, and the README says so louder than I'm saying it here.
Two deploy modes: a hosted engine (FastAPI on a HF Space) behind a Cloudflare Worker, or a fully client-side build compiled to WebAssembly via Pyodide — the document never leaves the browser tab in that mode.
304 tests across four suites, real CI on every push, no accounts (anonymous HMAC-token workspace).
Live:
https://watermark-finder.pages.dev/
Source:
https://github.com/Mati83mon/Watermark_Finder
Detects six distinct covert channels with exact character offsets, not guesses: Unicode tag payloads, variation-selector encoding, zero-width binary/markers, homoglyph substitution, bidi controls, and repetition stamps. Also verifies C2PA content credentials (Adobe, Leica, Google-signed manifests) — and keeps "integrity" (did the bytes change?) strictly separate from "trust" (is the signer who they claim?), because a forged certificate verifies perfectly too.
It scores writing register (how much prose reads like unedited assistant output) with visible uncertainty ranges and per-feature contributions — and refuses a verdict below ~15 words rather than fake precision. That score measures register, not authorship, and the README says so louder than I'm saying it here.
Two deploy modes: a hosted engine (FastAPI on a HF Space) behind a Cloudflare Worker, or a fully client-side build compiled to WebAssembly via Pyodide — the document never leaves the browser tab in that mode.
304 tests across four suites, real CI on every push, no accounts (anonymous HMAC-token workspace).
Live:
https://watermark-finder.pages.dev/
Source:
https://github.com/Mati83mon/Watermark_Finder