GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Sankar-2910
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Sankar-2910
Model Tree
Input Modalities
Output Modalities
Supported Functionality
Most public language models start from billions of pretrained weights.
This one doesn't.
Every one of its 26.11 million parameters began as random numbers and learned exclusively from a curated Gen Z → English translation dataset.
Built specifically for:
| Gen Z | Standard English |
|---|---|
no cap | I'm being completely honest. |
say less, I'm tryna vibe | I understand, I'm trying to relax. |
she's mogging everyone | She's outshining everyone by looking better. |
that's straight cap | That's an outright lie. |
after the argument she crashed out | After the argument she had an emotional outburst. |
The goal isn't to sound like ChatGPT.
The goal is to preserve meaning while removing slang naturally.
| Property | Value |
|---|---|
| Architecture | Llama-style Decoder-only Transformer |
| Parameters | 26.07M |
| Layers | 7 |
| Hidden Size | 448 |
| Attention Heads | 7 |
| Context Length | 384 |
| Vocabulary | 8,000 |
| Tokenizer | Custom Byte-Level BPE |
| Training | From Scratch |
The model is intentionally specialized.
Think of it as a translator, not a general assistant.
This is the project's biggest differentiator.
Instead of fine-tuning Llama, Qwen, or Mistral, the entire network was trained from random initialization.
That means:
text
Layers 7Hidden Size 448Attention Heads 7KV Heads 7Intermediate Size 1792Context Length 384Vocabulary 8000Activation SiLURoPE ✓RMSNorm ✓Tied Embeddings ✓
Although it follows the Llama architecture, it does not reuse Meta's pretrained weights.
The model learned from 139,074 cleaned instruction-response pairs.
| Split | Size |
|---|---|
| Train | 90% |
| Validation | 5% |
| Test | 5% |
The split was deterministic using a fixed seed and stratified by instruction type.
Two prompt styles were used consistently:
Example training prompt:
text
<s><|instruction|>Translate the following Gen Z slang sentence into clear, standard English.<|input|>bro that fit is so mid ngl<|response|>
Keeping one prompt format helped the model specialize instead of behaving like a chatbot.
The tokenizer wasn't borrowed either.
It was trained exclusively on the Gen Z dataset.
| Property | Value |
|---|---|
| Type | Byte-Level BPE |
| Vocabulary | 8,000 |
| Training Data | Only this dataset |
Because of this custom tokenizer, additional compatibility work was required for GGUF conversion and llama.cpp support.
Internal evaluation on 50 randomly sampled test examples showed:
| Metric | Score |
|---|---|
| Exact Match | 69% |
| Semantic Match | ~97% |
| Crash Stability | 100% |
The exact-match metric is intentionally strict.
Many "incorrect" outputs are actually valid paraphrases.
Example:
Input
say less
Output
I understand.
Different wording.
Same meaning.
| File | Purpose |
|---|---|
model.safetensors | Hugging Face model |
config.json | Model configuration |
tokenizer.json | Custom tokenizer |
genz-translator-f16.gguf | Full precision |
genz-translator-q8_0.gguf | Recommended GGUF |
genz-translator-q8_0.gguf
Best balance between quality and local CPU performance.
python
from transformers import AutoTokenizer, AutoModelForCausalLMimport torchrepo = "Sankar-2910/genz-translator"tokenizer = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo)prompt = ("<s><|instruction|>""Translate the following Gen Z slang sentence into clear, standard English.""<|input|>no cap""<|response|>")inputs = tokenizer(prompt, return_tensors="pt")with torch.no_grad():output = model.generate(**inputs,max_new_tokens=64,do_sample=False,temperature=0,eos_token_id=tokenizer.eos_token_id,pad_token_id=tokenizer.eos_token_id)print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:],skip_special_tokens=True))
Output:
text
I'm being completely honest.
bash
llama-cli \-m genz-translator-q8_0.gguf \-p "<s><|instruction|>Translate the following Gen Z slang sentence into clear, standard English.<|input|>bro that fit is so mid ngl<|response|>" \-n 96
Create a Modelfile.
text
FROM ./genz-translator-q8_0.ggufTEMPLATE """<s><|instruction|>Translate the following Gen Z slang sentence into clear, standard English.<|input|>{{ .Prompt }}<|response|>"""PARAMETER temperature 0PARAMETER num_predict 160PARAMETER stop "</s>"
Build:
bash
ollama create genz-translator -f Modelfile
Run:
bash
ollama run --raw genz-translator "bro that fit is so mid ngl"
Using --raw preserves the original training prompt format.
Use the same structure the model was trained on.
text
<s><|instruction|>Translate the following Gen Z slang sentence into clear, standard English.<|input|>no cap<|response|>
Avoid chat-style prompts.
The model is optimized for single-turn translation, not conversational memory.
This model was designed for lightweight local inference.
| Quantization | Approx. Memory |
|---|---|
| Q8_0 | ~30 MB |
| F16 | ~50 MB |
Recommended settings:
temperature=0num_predict≈160Being trained from scratch on a specialized dataset means:
The trade-off is specialization: small, fast, and purpose-built.
If you use this model in research, please cite:
bibtex
@misc{genztranslator2026,title={GenZ Translator},author={Sankar Narayanan},year={2026},note={26M parameter decoder-only transformer trained entirely from scratch for Gen Z slang translation},url={https://huggingface.co/Sankar-2910/genz-translator}}
If referencing the underlying architecture:
bibtex
@article{touvron2023llama,title={LLaMA: Open and Efficient Foundation Language Models},author={Touvron, Hugo and others},journal={arXiv preprint arXiv:2302.13971},year={2023}}
. This project was created by Sankar Narayanan as an exploration of training compact language models from scratch, custom tokenization, GGUF deployment, and local AI inference.
If you build something with it, I'd love to see it.
| Quantizations | F16 & Q8_0 |