Yoru

Yoru is a model in the Mori family. It is a full fp16 fine-tune of SmolLM2-1.7B-Instruct for casual internet / Gen Z chat.

This repo is the merged weights. Load it with Transformers only. No PEFT, no separate base download. It is a style / persona mix, not a facts model.

Model Details

• Name: Yoru
• Family: Mori
• Repo: grenishrai/yoru
• Base: HuggingFaceTB/SmolLM2-1.7B-Instruct
• Type: Causal LM, QLoRA SFT then merged to fp16
• Weights: model.safetensors (~3.42 GB, fp16)
• Language: English (internet slang)
• License: MIT (this checkpoint; the base model has its own license)
• Dataset: grenishrai/genz-sft-dataset (3,511 train rows)
• Kept checkpoint: epoch 2 (checkpoint-404), selected by lowest validation loss, then merged

Uses

Direct Use

• Chat replies in a short, informal online voice
• Style-transfer / persona experiments
• Side-by-side checks against the base Instruct model

Pass the same system prompt the data used, or the voice will slip.

Out-of-Scope Use

• Treating replies as real Gen Z speech or sociolinguistic ground truth
• Formal, clinical, legal, medical, or factual QA
• Impersonating a real person
• Safety-critical or production assistant without extra instruction / safety data

How to Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "grenishrai/yoru"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)
model.eval()

system = (
    "You are a pure Gen Z speaker. Always reply in natural Gen Z slang and internet speech. "
    "Use words and phrases like: no cap, fr fr, lowkey, highkey, bet, rizz, mid, slay, periodt, "
    "it's giving, sus, oof, vibes, down horrendous, bussin, cooked, goated, main character, etc. "
    "Keep replies casual, short to medium length, and online. Never break character. "
    "Never explain the slang. Never sound formal or like a normal AI."
)

messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": "I barely slept and now I have to be a person today"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    out = model.generate(
        inputs,
        max_new_tokens=80,
        do_sample=True,
        temperature=0.8,
        top_p=0.9,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

eos and pad are both <|im_end|>. That is how SmolLM2-Instruct is set up. ChatML template is in this repo.

Optional 4-bit load at inference:

from transformers import BitsAndBytesConfig

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)
model = AutoModelForCausalLM.from_pretrained(
    "grenishrai/yoru",
    quantization_config=bnb,
    device_map="auto",
)

Training Details

Training Data

grenishrai/genz-sft-dataset: synthetic 3-turn chats (system, user, assistant). One fixed system prompt. Everyday user turns. Informal assistant replies.

An 8% holdout (seed=42) was used for validation. There is no public test split.

Training Procedure

QLoRA SFT with TRL SFTTrainer on a Kaggle T4. Completion-only NLL. AMP off (fp16=False, bf16=False) because T4 + SmolLM2 bf16 LoRA hits a GradScaler crash. After training, the epoch-2 adapter was merged into the base weights and saved as this fp16 checkpoint.

Setting	Value
Base	HuggingFaceTB/SmolLM2-1.7B-Instruct
Train quantization	4-bit NF4, double quant, fp16 compute
Released weights	merged fp16
LoRA rank / alpha / dropout	32 / 64 / 0.05
LoRA targets	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Epochs run	3
Checkpoint kept	epoch 2 (best eval_loss)
Effective batch	16 (4 × 4 grad accum)
Learning rate	2e-4, cosine, ~5% warmup
Weight decay / grad clip	0.01 / 0.3
Max length	512
Optimizer	paged_adamw_8bit
Loss	completion-only NLL
Seed	42

Earlier Colab runs on the same data (4 and 9 epochs, same LoRA) also peaked at epoch 2. Extra epochs dropped train loss and raised val loss. Do not train this mix much past 2–3 epochs.

Evaluation

Judge this model by generations, not token accuracy. Open-ended slang chat will sit near 56% mean token accuracy even when the voice has moved.

Kaggle 3-epoch run (the checkpoint merged into this repo):

Epoch	Train loss	Val loss	Mean token acc
1	2.30	2.21	53%
2	1.89	2.05	56%
3	1.41	2.12	56%

Best checkpoint: epoch 2. Epoch 3 was worse on the holdout.

Informal demo prompts after that run (not a benchmark):

Prompt	Yoru (approx.)	Note
I barely slept and now I have to be a person today	short slang, still a pep line	OK
Should I text first or wait	said wait	Wrong take vs the data (send one normal text)
My boss emailed me after hours	don’t answer, then some garble	Right instinct
Rate this meal: leftover rice and hot sauce	W. staple not a flop	Best of the four

Vs the base Instruct model the voice is a clear step: less fake hype and emoji salad. It is not a locked persona. The dataset mixes lowercase chat (65%) and ALL-CAPS mishap captions (35%), so generations can mix those styles.

Bias, Risks, and Limitations

• Slang here is a stylized internet register, not a sample of any age group or region.
• Slang dates quickly. Some lines will read as forced.
• Replies can be blunt or dismissive. That is the target style, not a general assistant policy.
• The model can mash slang without a clear take, or miss the intended advice (see “text first”).
• 1.7B + 3.5k synthetic rows will not match a larger chat model on reasoning or facts.
• No safety alignment beyond whatever the base Instruct model already has.

Recommendations

• Always send the system prompt above.
• Prefer val loss + human reads over token accuracy.
• Mix with general instruction / safety data if you ship this.
• For a single voice, clean the ALL-CAPS caption rows in the dataset and train again. More epochs will not fix that.

Environmental Impact

• Hardware: Kaggle Tesla T4
• Hours used: about 45–70 minutes for the 3-epoch run
• Cloud provider: Kaggle

Technical Specs

• Architecture: SmolLM2 1.7B Instruct (Llama-style), 24 layers, hidden 2048
• Objective: next-token SFT on assistant tokens only
• Files in this repo: model.safetensors, config.json, generation_config.json, tokenizer, ChatML template
• Not in this repo: LoRA adapter files (removed after merge)

Citation

@misc{yoru-2026,
  title  = {Yoru (Mori): SmolLM2-1.7B-Instruct fine-tune for casual internet chat},
  author = {grenishrai},
  year   = {2026},
  url    = {https://huggingface.co/grenishrai/yoru}
}

Model Card Contact

Open an issue on the model repository.