Liquid-Thinking-Preview
This repository contains a full-parameter fine-tuned version of LiquidAI/LFM2-1.2B on the HelpingAI/Intermediate-Thinking-130k dataset. The goal of this fine-tune is to enhance the model's ability to perform step-by-step reasoning and solve problems that require intermediate thought processes.
This model was trained with high efficiency using the Unsloth library.
- GGUF Quantizations: For CPU inference, GGUF versions are available in the GGUF repository here.
Model Details
- Base Model: LiquidAI/LFM2-1.2B
- Dataset: HelpingAI/Intermediate-Thinking-130k
- Fine-tuning Method: Full Parameter Fine-tuning.
- Frameworks: Unsloth, Transformers, PyTorch, TRL.
How to Use with Unsloth
You can easily run this model for fast inference using the unsloth library.
First, install the necessary packages:
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install --no-deps git+https://github.com/huggingface/transformers.git
pip install --no-deps causal-conv1d==1.5.0.post8
Next, use the following Python code to load the model and generate text:
from unsloth import FastModel
import torch
from transformers import TextStreamer
model_repo_id = "kreasof-ai/Liquid-Thinking-Preview"
model, tokenizer = FastModel.from_pretrained(
model_name = model_repo_id,
dtype = None, # None for auto detection
load_in_4bit = True, # Use 4bit quantization for faster inference
)
# Define the conversation
messages = [{
"role": "user",
"content": "Explain the step-by-step process of photosynthesis in simple terms.",
}]
# Apply the chat template
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt = True, # Must add for generation
return_tensors = "pt",
).to("cuda")
# Set up the streamer for real-time output
streamer = TextStreamer(tokenizer, skip_prompt = True)
# Generate the response
_ = model.generate(
**inputs,
max_new_tokens = 4096,
# Recommended generation parameters from LiquidAI
temperature = 0.3,
min_p = 0.15,
repetition_penalty = 1.05,
use_cache = True,
streamer = streamer,
)
Training Procedure
Data Preprocessing
The HelpingAI/Intermediate-Thinking-130k dataset, consisting of instruction and output pairs, was formatted into a conversational format using the LFM2's chat template. The template follows this structure:
<|im_start|>user
{instruction}<|im_end|>
<|im_start|>assistant
{output}<|im_end|>
The training loss was calculated only on the assistant's responses (train_on_responses_only), encouraging the model to learn the desired output format and reasoning style without being penalized for the user's prompt.
Hyperparameters
The model was trained with the following hyperparameters using trl.SFTTrainer:
| Hyperparameter | Value |
|---|---|
per_device_train_batch_size | 8 |
gradient_accumulation_steps | 4 |
Effective Batch Size | 32 |
num_train_epochs | 3 |
learning_rate | 4e-5 |
lr_scheduler_type | cosine |
warmup_ratio | 0.1 |
optim | adamw_torch |
weight_decay | 0.01 |
max_seq_length | 8192 |
seed | 3407 |
Training Results
The training loss curve showed a steady decrease, indicating successful learning.
Intended Use & Limitations
This model is intended for research and experimentation in multi-step reasoning, complex problem-solving, and chain-of-thought generation.
- Limitations: Like all language models, it is prone to hallucination and may produce factually incorrect or biased information.
- Out-of-Scope: This model is not designed for production use without extensive testing and safety evaluations. It should not be used for critical applications where incorrect information could cause harm.