metadata

language:
  - ru
datasets:
  - IlyaGusev/saiga_scored
  - IlyaGusev/saiga_preferences
license: gemma

QuantFactory/saiga_gemma2_9b-GGUF

This is quantized version of IlyaGusev/saiga_gemma2_9b created using llama.cpp

Original Model Card

Saiga/Gemma2 9B, Russian Gemma-2-based chatbot

Based on Gemma-2 9B Instruct.

Prompt format

Gemma-2 prompt format:

<start_of_turn>system
Ты — Сайга, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им.<end_of_turn>
<start_of_turn>user
Как дела?<end_of_turn>
<start_of_turn>model
Отлично, а у тебя?<end_of_turn>
<start_of_turn>user
Шикарно. Как пройти в библиотеку?<end_of_turn>
<start_of_turn>model

Code example

# Исключительно ознакомительный пример.
# НЕ НАДО ТАК ИНФЕРИТЬ МОДЕЛЬ В ПРОДЕ.
# См. https://github.com/vllm-project/vllm или https://github.com/huggingface/text-generation-inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig

MODEL_NAME = "IlyaGusev/saiga_gemma2_10b"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    load_in_8bit=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
generation_config = GenerationConfig.from_pretrained(MODEL_NAME)
print(generation_config)

inputs = ["Почему трава зеленая?", "Сочини длинный рассказ, обязательно упоминая следующие объекты. Дано: Таня, мяч"]
for query in inputs:
    prompt = tokenizer.apply_chat_template([{
        "role": "user",
        "content": query
    }], tokenize=False, add_generation_prompt=True)
    data = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
    data = {k: v.to(model.device) for k, v in data.items()}
    output_ids = model.generate(**data, generation_config=generation_config)[0]
    output_ids = output_ids[len(data["input_ids"][0]):]
    output = tokenizer.decode(output_ids, skip_special_tokens=True).strip()
    print(query)
    print(output)
    print()
    print("==============================")
    print()

Versions

v2:

258869abdf95aca1658b069bcff69ea6d2299e7f
Other name: saiga_gemma2_9b_abliterated_sft_m3_d9_abliterated_kto_m1_d13
SFT dataset config: sft_d9.json
SFT model config: saiga_gemma2_9b_sft_m2.json
KTO dataset config: pref_d11.json
KTO model config: saiga_gemma2_9b_kto_m1.json
SFT wandb: link
KTO wandb: link

v1:

fa63cfe898ee6372419b8e38d35f4c41756d2c22
Other name: saiga_gemma2_9b_abliterated_sft_m2_d9_abliterated_kto_m1_d11
SFT dataset config: sft_d9.json
SFT model config: saiga_gemma2_9b_sft_m2.json
KTO dataset config: pref_d11.json
KTO model config: saiga_gemma2_9b_kto_m1.json
SFT wandb: link
KTO wandb: link

Evaluation

Dataset: https://github.com/IlyaGusev/rulm/blob/master/self_instruct/data/tasks.jsonl
Framework: https://github.com/tatsu-lab/alpaca_eval
Evaluator: alpaca_eval_cot_gpt4_turbo_fn

Pivot: gemma_2_9b_it_abliterated

model	length_controlled_winrate	win_rate	standard_error	avg_length
gemma_2_9b_it_abliterated	50.00	50.00	0.00	1126
saiga_gemma2_9b, v1	48.66	45.54	2.45	1066
saiga_gemms2_9b, v2	47.77	45.30	2.45	1074