Rewrites formal Russian into casual chat-style Russian (Telegram-like), without profanity, while keeping facts, names, numbers, and paragraph structure.

Not a general chat model — use the instruction prefix from training (see below).

Model lineage

Stage	Artifact
Base	google/gemma-4-E4B-it
LoRA (SFT)	pavelfedortsov/gemma4-e4b-lora-colloquial-ru
This repo	LoRA merged into base + vLLM fixes (`k_norm`, processor configs)

Merge was done with peft.merge_and_unload(); missing language_model k_norm weights for layers 24–41 were copied from the base checkpoint (required for vLLM).

Training data

50,000 SFT pairs, mat-free colloquial style
Hub dataset: pavelfedortsov/russian-colloquial-sft-50k
Built from kurumikz/telegram-corpus-russian-kazakh + Gemini pair generation (see dataset card)

User prompt template (training & inference):

Перепиши простым разговорным русским, как в переписке. Без мата и грубости. Сохрани смысл:
<формальный текст>

Training configuration (LoRA → merge)

Config file (also in card_assets/train_colloquial_e4b_gpu.yaml):

Parameter	Value
Base model	`google/gemma-4-E4B-it`
Method	LoRA on language tower (`model.language_model.*`)
LoRA rank / alpha	32 / 64
Target modules	`q,k,v,o` + MLP (`gate, up, down`)
Dataset	50k × 1 repeat
Epochs	2 (12,500 optimizer steps)
Seq length	512
Batch	1 × grad accum 8 (effective 8)
LR	1e-4, cosine, warmup 3%
Precision	bf16, gradient checkpointing
Loss	assistant-only
Hardware	RunPod A100 80GB

Training metrics (LoRA run)

Metric	Start (step ~25)	End (step 12,500)	Best
`loss`	~3.42	~0.81	~0.67
`mean_token_accuracy`	~0.63	~0.82	~0.84

Checkpoints saved every 1000 steps under the LoRA adapter repo.

Inference

RunPod Serverless (vLLM)

env

MODEL_NAME=pavelfedortsov/gemma4-e4b-colloquial-ru-merged
HF_TOKEN=<your_token>
TRUST_REMOTE_CODE=true
DTYPE=bfloat16
MAX_MODEL_LEN=4096
GPU_MEMORY_UTILIZATION=0.90
ENFORCE_EAGER=true
ENABLE_LORA=false
LANGUAGE_MODEL_ONLY=true
LIMIT_MM_PER_PROMPT={"image":0,"audio":0,"video":0}

Recommended GPU: ≥40 GB VRAM (merged ~32 GB weights in bf16).

Transformers (local)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "pavelfedortsov/gemma4-e4b-colloquial-ru-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

formal = "Сегодня на совещании обсуждали внедрение новой версии API."
user = (
    "Перепиши простым разговорным русским, как в переписке. "
    "Без мата и грубости. Сохрани смысл:\n"
    f"{formal}"
)
messages = [{"role": "user", "content": user}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.9)
print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

OpenAI-compatible API (RunPod / vLLM)

bash

curl "$RUNPOD_URL/v1/chat/completions" \
  -H "Authorization: Bearer $RUNPOD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pavelfedortsov/gemma4-e4b-colloquial-ru-merged",
    "messages": [{
      "role": "user",
      "content": "Перепиши простым разговорным русским, как в переписке. Без мата и грубости. Сохрани смысл:\nВаш формальный текст."
    }],
    "max_tokens": 512,
    "temperature": 0.7
  }'

Limitations

Gemma license applies to the base architecture and weights.
Quality varies on long news-style text; model may shorten or paraphrase aggressively.
Not safety-tuned for production without your own evaluation.
Merged vs LoRA inference can differ slightly in style.

Related repos

Resource	Link
LoRA adapter	https://huggingface.co/pavelfedortsov/gemma4-e4b-lora-colloquial-ru
Dataset (50k)	https://huggingface.co/datasets/pavelfedortsov/russian-colloquial-sft-50k
Base model	https://huggingface.co/google/gemma-4-E4B-it