LiberalMind Γ— Spark-X2.6

A parameter-efficient adaptation of Spark-X2.6.

This repository contains a LiberalMind LoRA adapter designed to extend the capabilities of the Spark-X2.6 language model.

Instead of distributing a complete language model checkpoint, this project provides a lightweight set of trainable adapter weights that can be attached to the original Spark-X2.6 architecture.

The base model provides the underlying language capabilities, while the LoRA adapter introduces additional learned parameters.

Developed and published by LiberalMind.


Model Information

Property Value
Organization LiberalMind
Base model family Spark-X2.6
Base model publisher XHToken
Model size 4B (if using the 4B variant)
Fine-tuning method LoRA
Adapter framework Hugging Face PEFT
Inference framework Transformers
Model format PEFT adapter
Training approach Parameter-efficient fine-tuning

Base Model

Official model repositories:

Important: The adapter must be loaded onto the exact base checkpoint used during its original training. The base_model_name_or_path field in adapter_config.json identifies the intended base model. Do not substitute the Base or post-trained variant without checking compatibility.


1. How It Works

LoRA (Low-Rank Adaptation) fine-tunes a relatively small number of additional parameters while keeping the original model weights frozen.

Conceptually:

             Original Spark-X2.6
                      |
                      |
                Frozen Weights
                      |
                      |
             +----------------+
             |                |
             v                v
        Original Layer    LoRA Adapter
             |                |
             |                |
             +-------+--------+
                     |
                     v
              Combined Output
                     |
                     v
             Adapted Language Model

The adapter stores learned low-rank updates rather than another copy of the entire model.

This makes it possible to distribute and share fine-tuned adaptations without redistributing the full base checkpoint.


2. Repository Structure

Expected files:

spark-x2.6-lora/
β”‚
β”œβ”€β”€ README.md
β”œβ”€β”€ adapter_config.json
β”œβ”€β”€ adapter_model.safetensors
β”‚
└── tokenizer files (optional)
File Purpose
adapter_config.json LoRA architecture and base model metadata
adapter_model.safetensors Learned LoRA parameters
README.md Documentation

The original Spark-X2.6 weights are downloaded separately.


3. Installation

Recommended environment:

  • Python 3.10+
  • PyTorch
  • Transformers
  • PEFT
  • Accelerate
  • NVIDIA GPU for the example training workflow

Install dependencies:

pip install torch
pip install transformers==4.57.1
pip install peft accelerate safetensors
pip install datasets

The Transformers version follows the official Spark-X2.6 integration example.


4. Download and Load the Adapter

You do not need to manually merge LoRA tensors into the base model.

PEFT handles the adapter integration.

import torch

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer
)

from peft import (
    PeftConfig,
    PeftModel
)

# LiberalMind adapter
ADAPTER_ID = "LiberalMind/spark-x2.6-lora"

# Read the original base model identifier from the adapter
peft_config = PeftConfig.from_pretrained(ADAPTER_ID)

BASE_MODEL = peft_config.base_model_name_or_path

print("Base model:", BASE_MODEL)
print("LoRA configuration:", peft_config)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL,
    trust_remote_code=True
)

# Load original Spark-X2.6
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto"
)

# Attach LiberalMind LoRA weights
model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID
)

model.eval()

print("LiberalMind adapter successfully loaded!")

If the adapter metadata contains an obsolete or local base-model path, replace BASE_MODEL manually with the exact compatible Hugging Face model ID.


5. Run Inference

Once the adapter is attached, the resulting model can be used through the regular Transformers generation interface.

import torch

messages = [
    {
        "role": "user",
        "content": "Explain how a neural network learns."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )

generated = outputs[0][inputs["input_ids"].shape[-1]:]

print(
    tokenizer.decode(
        generated,
        skip_special_tokens=True
    )
)

The chat template is important: it formats conversations according to Spark-X2.6's expected input structure.


6. Continue Training the Existing LoRA Adapter

An existing adapter can be loaded in trainable mode and fine-tuned further.

Unlike inference mode, this enables gradients for the adapter parameters.

Step 1. Load the base model and existing adapter

import torch

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

ADAPTER_ID = "LiberalMind/spark-x2.6-lora"

peft_config = PeftConfig.from_pretrained(ADAPTER_ID)
BASE_MODEL = peft_config.base_model_name_or_path

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL,
    trust_remote_code=True
)

dtype = (
    torch.bfloat16
    if torch.cuda.is_available()
    and torch.cuda.is_bf16_supported()
    else torch.float16
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    trust_remote_code=True,
    torch_dtype=dtype,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
    is_trainable=True
)

model.config.use_cache = False

print("Loaded existing LoRA adapter in training mode.")

trainable = sum(
    p.numel() for p in model.parameters() if p.requires_grad
)

total = sum(
    p.numel() for p in model.parameters()
)

print(f"Trainable parameters: {trainable:,}")
print(f"Total parameters: {total:,}")
print(f"Trainable percentage: {100 * trainable / total:.4f}%")

Step 2. Prepare the dataset

The following example expects a JSONL file:

train.jsonl

Example records:

{"messages":[{"role":"user","content":"What is machine learning?"},{"role":"assistant","content":"Machine learning is a field of AI that enables systems to learn patterns from data."}]}
{"messages":[{"role":"user","content":"Explain gradient descent."},{"role":"assistant","content":"Gradient descent is an optimization algorithm that iteratively adjusts parameters to minimize a loss function."}]}

Each record contains a conversation with user and assistant messages.

Step 3. Prepare the training dataset

from datasets import load_dataset
from transformers import DataCollatorForLanguageModeling

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

dataset = load_dataset(
    "json",
    data_files="train.jsonl",
    split="train"
)

def format_example(example):
    text = tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False,
        enable_thinking=True
    )
    return {"text": text}

dataset = dataset.map(format_example)

def tokenize_function(examples):
    return tokenizer(
        examples["text"],
        truncation=True,
        max_length=4096
    )

tokenized_dataset = dataset.map(
    tokenize_function,
    batched=True,
    remove_columns=dataset.column_names
)

collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=False
)

This is a basic full-sequence causal language-modeling example. It trains on both user and assistant tokens; for production instruction tuning, assistant-only loss masking may be preferable.

Step 4. Start training

import torch

from transformers import (
    Trainer,
    TrainingArguments
)

OUTPUT_DIR = "./liberalmind-spark-lora-trained"

bf16 = (
    torch.cuda.is_available()
    and torch.cuda.is_bf16_supported()
)

fp16 = (
    torch.cuda.is_available()
    and not bf16
)

training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,

    num_train_epochs=1,

    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,

    learning_rate=1e-5,
    warmup_ratio=0.03,
    lr_scheduler_type="cosine",

    logging_steps=10,

    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,

    bf16=bf16,
    fp16=fp16,

    gradient_checkpointing=True,

    max_grad_norm=1.0,

    report_to="none",
    remove_unused_columns=False
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    data_collator=collator
)

trainer.train()

trainer.save_model(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)

print("Training completed!")
print("Adapter saved to:", OUTPUT_DIR)

This workflow loads the existing adapter and updates its parameters. It does not train a new independent LoRA adapter from scratch.

To resume optimizer and scheduler states from an interrupted Trainer run, use its saved checkpoint-* directory:

trainer.train(
    resume_from_checkpoint="./liber almind-spark-lora-trained/checkpoint-200"
)

For that exact resume, the checkpoint must contain the Trainer state, not merely the final adapter weights.


7. Export the Adapted Model

The adapter can optionally be merged into the base model.

merged_model = model.merge_and_unload()

merged_model.save_pretrained(
    "./liber almind-spark-merged",
    safe_serialization=True
)

tokenizer.save_pretrained(
    "./liber almind-spark-merged"
)

print("Merged model exported.")

The merged checkpoint contains the combined model weights rather than a separate LoRA adapter.

The resulting model is substantially larger than the adapter-only repository.


8. Training and Reproducibility

For reproducibility, the following should be preserved:

  • Original base-model identifier.
  • adapter_config.json.
  • LoRA adapter weights.
  • Training dataset and preprocessing.
  • Learning rate and batch configuration.
  • Training framework versions.
  • Training checkpoints, when exact optimizer-state recovery is required.

A LoRA checkpoint by itself does not contain the full base language model.

Anyone using this repository must have access to the corresponding original Spark-X2.6 checkpoint.


9. Research Direction

LiberalMind explores lightweight model adaptation and efficient AI development.

This project represents an approach to extending existing language model capabilities without requiring a complete new pretraining run.

Potential future directions include:

  • Further instruction tuning.
  • Mathematical reasoning.
  • Coding and software engineering.
  • Tool-use and agent workflows.
  • Model efficiency and deployment.
  • Evaluation across different reasoning and problem-solving tasks.

No benchmark improvement is claimed without a corresponding evaluation.


Credits

Base architecture and pretrained weights: XHToken / SparkLLM

Adapter development: LiberalMind

Fine-tuning technology: Hugging Face PEFT

Inference framework: Transformers

License: The base model is published under Apache 2.0. Verify the adapter's own licensing and attribution requirements before redistribution.


LiberalMind

Building intelligence beyond the screen.

References

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support