🦁 Russian IT Community LoRA Model Zoo

58 pre-trained adapters (55 domain adapters + 3 flagship 7B–8B QLoRA), fine-tuned on the RICC corpus (2.91M messages, 171.5k curated SFT dialogues) for Russian-language IT discourse: backend, DevOps, AI/ML, infrastructure.

Catalog regenerated from the Hub file tree on 2026-09-02. Source of truth: the siblings listing of this repository.

⚡ Quick Start: 3-Line Inference

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen/Qwen2.5-1.5B-Instruct"  # any base model from the catalog
adapter_id = "wwewtech/russian-it-community-lora"
subfolder = "qwen2.5_1.5b_instruct"  # choose from the catalog below

tokenizer = AutoTokenizer.from_pretrained(model_id)
base_model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_id, subfolder=subfolder)

inputs = tokenizer("<|user|>\nКак настроить репликацию PostgreSQL?\n<|assistant|>\n", return_tensors="pt").to(
    model.device
)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📚 Full Catalog (58 Adapters)

🥇 Flagship QLoRA Models (7B–8B)

Full-precision copies also live under models/: models/heavyweight_qwen2.5_coder_7b, models/heavyweight_deepseek_r1_7b, models/heavyweight_llama3.1_8b.

📓 Training Data & Evaluation Status

• Training corpus: RICC SFT Dialogues (171,520 multi-turn dialogues).
• Training regime: Published adapters are domain-adaptation pilot checkpoints (50–100 training steps on sampled domain batches) demonstrating PEFT compatibility and consumer-GPU deployment (RTX 3060 12GB). They are not multi-epoch checkpoints over the entire corpus.
• Academic benchmark numbers (HumanEval / RuMMLU / PPL) published earlier are withdrawn pending re-evaluation: the harness had answer-parsing and column-mapping defects that produced implausible values (see repo commit history). Enterprise scenario scores are rubric-based heuristics, not capability measurements.