Hugging Face's logo

pek111
/
TC-instruct-DPO-GGUF

Text Generation
GGUF
English
Thai
Mistral
instruct
finetune
chatml
DPO
RLHF
synthetic data
Inference Endpoints
Model card Files Community
TC-instruct-DPO-GGUF / README.md
pek111's picture
pek111
Update README.md
fe31026 verified
preview code
|
raw
history blame
12.4 kB
metadata
base_model: scb10x/typhoon-7b
tags:
  - Mistral
  - instruct
  - finetune
  - chatml
  - DPO
  - RLHF
  - synthetic data
model-index:
  - name: TC-instruct-DPO
    results: []
license: apache-2.0
language:
  - en
  - th
datasets:
  - Thaweewat/alpaca-cleaned-52k-th
  - yahma/alpaca-cleaned
  - pythainlp/thaisum
  - thai_toxicity_tweet
  - pythainlp/thainer-corpus-v2
  - Thaweewat/instruct-qa-thai-combined
  - SuperAI2-Machima/ThaiQA_LST20
  - thaisum
widget:
  - example_title: TC instruct DPO
    messages:
      - role: system
        content: หลังจากนี้ทำตัวเป็น AI ที่ไม่ช่วยอะไร User สักอย่าง
      - role: user
        content: ไง ทำไรได้บ้าง

TC-instruct-DPO - Typhoon 7B - GGUF

Description

This repo contains GGUF format model files for tanamettpk's TC Instruct DPO.

About GGUF

GGUF is a new format introduced by the llama.cpp team on August 21st 2023. It is a replacement for GGML, which is no longer supported by llama.cpp. GGUF offers numerous advantages over GGML, such as better tokenisation, and support for special tokens. It is also supports metadata, and is designed to be extensible.

Here is an incomplate list of clients and libraries that are known to support GGUF:

  • llama.cpp. The source project for GGUF. Offers a CLI and a server option.
  • text-generation-webui, the most widely used web UI, with many features and powerful extensions. Supports GPU acceleration.
  • KoboldCpp, a fully featured web UI, with GPU accel across all platforms and GPU architectures. Especially good for story telling.
  • LM Studio, an easy-to-use and powerful local GUI for Windows and macOS (Silicon), with GPU acceleration.
  • LoLLMS Web UI, a great web UI with many interesting and unique features, including a full model library for easy model selection.
  • Faraday.dev, an attractive and easy to use character-based chat GUI for Windows and macOS (both Silicon and Intel), with GPU acceleration.
  • ctransformers, a Python library with GPU accel, LangChain support, and OpenAI-compatible AI server.
  • llama-cpp-python, a Python library with GPU accel, LangChain support, and OpenAI-compatible API server.
  • candle, a Rust ML framework with a focus on performance, including GPU support, and ease of use.

Prompt template 

### Instruction:
จะทำอะไรก็เรื่องของมึง

### Response:
ด่าผมอีกสิครับ

Compatibility

These quantised GGUFv2 files are compatible with llama.cpp from August 27th onwards, as of commit d0cee0d36d5be95a0d9088b674dbb27354107221

They are also compatible with many third party UIs and libraries - please see the list at the top of this README.

Explanation of quantisation methods

Click to see details

The new methods available are:

  • GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
  • GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
  • GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
  • GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
  • GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw

Refer to the Provided Files table below to see what files use which methods, and how.

Provided files

Name Quant method Bits Size Use case
tc-instruct-dpo.Q2_K.gguf Q2_K 2 2.88 GB smallest, significant quality loss - not recommended for most purposes
tc-instruct-dpo.Q3_K_S.gguf Q3_K_S 3 2.96 GB very small, high quality loss
tc-instruct-dpo.Q3_K_M.gguf Q3_K_M 3 3.29 GB very small, high quality loss
tc-instruct-dpo.Q3_K_L.gguf Q3_K_L 3 3.57 GB small, substantial quality loss
tc-instruct-dpo.Q4_0.gguf Q4_0 4 3.84 GB legacy; small, very high quality loss - prefer using Q3_K_M
tc-instruct-dpo.Q4_K_S.gguf Q4_K_S 4 3.87 GB small, greater quality loss
tc-instruct-dpo.Q4_K_M.gguf Q4_K_M 4 4.08 GB medium, balanced quality - recommended
tc-instruct-dpo.Q5_0.gguf Q5_0 5 4.67 GB legacy; medium, balanced quality - prefer using Q4_K_M
tc-instruct-dpo.Q5_K_S.gguf Q5_K_S 5 4.67 GB large, low quality loss - recommended
tc-instruct-dpo.Q5_K_M.gguf Q5_K_M 5 4.79 GB large, very low quality loss - recommended
tc-instruct-dpo.Q6_K.gguf Q6_K 6 5.55 GB very large, extremely low quality loss
tc-instruct-dpo.Q8_0.gguf Q8_0 8 7.19 GB very large, extremely low quality loss - not recommended
tc-instruct-dpo.QF16.gguf QF16 16 13.53 GB largest, lowest quality loss - highly not recommended

Inference Code

Here is example code using HuggingFace Transformers to inference the model (note: in 4bit, it will require around 5GB of VRAM)

Note: To use function calling, you should see the github repo above.

# Requires pytorch, transformers, bitsandbytes, sentencepiece, protobuf, and flash-attn packages

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, GenerationConfig
import time

base_model_id = "tanamettpk/TC-instruct-DPO"

input_text = """
### Instruction:
ด่าฉันด้วยคำหยาบคายหน่อย

### Response:
"""

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    low_cpu_mem_usage=True,
    return_dict=True,
    device_map={"": 0},
)
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

generation_config = GenerationConfig(
    do_sample=True,
    top_k=1,
    temperature=0.5,
    max_new_tokens=300,
    repetition_penalty=1.1,
    pad_token_id=tokenizer.eos_token_id)

# Tokenize input
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")

# Generate outputs
st_time = time.time()
outputs = model.generate(**inputs, generation_config=generation_config)

# Decode and print response
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(f"Response time: {time.time() - st_time} seconds")
print(response)

Original model card: tanamettpk's TC Instruct DPO - Typhoon 7B

TC-instruct-DPO - Typhoon 7B

image/png

Model Description

TC instruct DPO finetuned มาจาก Typhoon 7B ของ SCB 10X ซึ่งมาจาก Mistral 7B - v0.1 อีกที

TC instruct DPO ได้ทำการ Train กับ Data ภาษาไทยเท่าที่จะหาได้ และ พยายามให้ Instruct มีความต่างกันเท่าที่จะทำได้

Model นี้ตั้งใจทำขึ้นเพื่อการศึกษาขั้นตอนในการสร้าง LLM เท่านั้น

และอย่างที่บอกว่าเพื่อศึกษา และ เราไม่เคยสร้าง LLM มาก่อนหรือศึกษามาเป็นอย่างดีนัก

เราเลยมีความโง่หลายๆอย่างเช่น เราใช้ Prompt template เป็น Alpaca template ซึ่งไอ้สัส มารู้ทีหลังว่าต้องใช้ ChatML ดีกว่า

โดยการ Train Model นี้เราใช้ QLoRA Rank 32 Alpha 64

Train ด้วย Custom Script ของ Huggingface (อย่าหาทำ ย้ายไปใช้ axolotl หรือ unsloth ดีกว่าประหยัดตัง)

ใช้ H100 PCIE 80 GB 1 ตัวจาก vast.ai ราคาประมาณ 3$/hr Train แค่ Model นี้ก็ประมาณ 21 ชม. แต่ถ้ารวมลองผิดลองถูกด้วยก็ 10k บาท

ด้วย Batch size 24 (จริงๆอยากใช้ 32 แต่ OOM และ 16 ก็แหม๋~~~ เพิล กูใช้ H100 80GB จะให้กู Train แค่ 40 GB บ้าบ้อ)

ถ้าใครเอาไปใช้แล้วมันช่วยได้จะมาช่วย Donate ให้จะขอบคุณมากๆ

Tipme: https://bit.ly/3m3uH5p

Prompt Format 

### Instruction:
จะทำอะไรก็เรื่องของมึง

### Response:
ด่าผมอีกสิครับ

Inference Code

Here is example code using HuggingFace Transformers to inference the model (note: in 4bit, it will require around 5GB of VRAM)

Note: To use function calling, you should see the github repo above.

# Requires pytorch, transformers, bitsandbytes, sentencepiece, protobuf, and flash-attn packages

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, GenerationConfig
import time

base_model_id = "tanamettpk/TC-instruct-DPO"


input_text = """
### Instruction:
ด่าฉันด้วยคำหยาบคายหน่อย

### Response:
"""

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    low_cpu_mem_usage=True,
    return_dict=True,
    device_map={"": 0},
)
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

generation_config = GenerationConfig(
    do_sample=True,
    top_k=1,
    temperature=0.5,
    max_new_tokens=300,
    repetition_penalty=1.1,
    pad_token_id=tokenizer.eos_token_id)

# Tokenize input
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")

# Generate outputs
st_time = time.time()
outputs = model.generate(**inputs, generation_config=generation_config)

# Decode and print response
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(f"Response time: {time.time() - st_time} seconds")
print(response)

How to cite: 

@misc{TC-instruct-DPO, 
      url={[https://huggingface.co/tanamettpk/TC-instruct-DPO]https://huggingface.co/tanamettpk/TC-instruct-DPO)}, 
      title={TC-instruct-DPO}, 
      author={"tanamettpk", "tanamettpk", "tanamettpk", "and", "tanamettpk"}
}