tahamajs/Qwen3-4b-gsm8k-Qlora-SFT

Model Card for outputs_sft

outputs_sft is a Supervised Fine-Tuning (SFT) LoRA adapter on top of [Qwen/Qwen3-4B]. It was trained with TRL and PEFT.

Quick start

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from peft import AutoPeftModelForCausalLM

REPO_ID = "outputs_sft"  # Replace with your Hub repo if different

# Load base model + merge LoRA on the fly (recommended for inference)
model = AutoPeftModelForCausalLM.from_pretrained(REPO_ID, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(REPO_ID, use_fast=True)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device_map="auto")

prompt = "If you had a time machine and could go only once, where and when would you go? Explain your reasoning."
out = pipe(prompt, max_new_tokens=256, do_sample=True, top_p=0.9, temperature=0.7)[0]["generated_text"]
print(out)

# Alternatively, if you already merged the LoRA and saved the full model weights:
# model = AutoModelForCausalLM.from_pretrained(REPO_ID, device_map="auto")
# tokenizer = AutoTokenizer.from_pretrained(REPO_ID, use_fast=True)

Intended uses & limitations

Intended uses

• General instruction following and helpful assistant style responses.
• Short-form reasoning and everyday Q&A.
• Creative writing, drafting, and rewriting.

Limitations

• Not evaluated for safety-critical or high-stakes domains.
• May produce inaccurate, biased, or undesired content.
• Long-chain reasoning may require specialized training.

Bias, risks, and limitations: Outputs may reflect biases present in training data. Review before use in production.

Training data

• Dataset not auto-detected from the notebook. Please document your data sources.

Training procedure

This model was trained with SFT using TRL/PEFT.

PEFT / LoRA Config

• lora_dropout: 0.05
• target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]

Precision & Quantization

• load_in_4bit: True
• bnb_4bit_compute_dtype: float32
• dtype: float32

Key hyperparameters

• num_train_epochs: 2
• per_device_train_batch_size: 10
• per_device_eval_batch_size: 10
• gradient_accumulation_steps: 2
• learning_rate: 9e-4
• lr_scheduler_type: cosine
• logging_steps: 2
• save_steps: 8
• save_strategy: steps
• bf16: True
• fp16: False
• seed: 42

Hardware & runtime

• GPU not detected from notebook logs.

Framework versions

• PEFT: 0.17.0
• TRL: 0.21.0
• Transformers: 4.55.1
• PyTorch: 2.8.0
• Datasets: 3.6.0
• Tokenizers: 0.21.4

Example prompts

• Explain diffusion models to a 12-year-old.
• Write a polite email asking for an extension on a project.
• Summarize the following text in 3 bullet points: ...

Evaluation

No formal evaluation metrics were logged in the notebook. If you run evaluations (e.g., on MT-Bench, MMLU, or a domain-specific set), please add the results here in a Model Index block or a table.

Pushing to the Hub

from huggingface_hub import HfApi, create_repo, upload_folder

REPO_ID = "outputs_sft"  # e.g., "YourUsername/outputs_sft"

# 1) Create the repo (once)
# create_repo(REPO_ID, repo_type="model", private=False)

# 2) Upload your adapter or merged model folder
upload_folder(
    repo_id=REPO_ID,
    folder_path="./outputs_sft",  # change to your output dir
    commit_message="Add SFT model",
)

License: Set license in the YAML header to a license compatible with the base model and your data (e.g., apache-2.0, mit, or the specific Qwen license if required).

Citations

@misc{vonwerra2022trl,
  title        = {{TRL: Transformer Reinforcement Learning}},
  author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
  year         = 2020,
  journal      = {GitHub repository},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/huggingface/trl}}
}

@misc{mukherjee2023peft,
  title        = {PEFT: Parameter-Efficient Fine-Tuning},
  author       = {Edward Hu and others},
  year         = 2023,
  howpublished = {\url{https://github.com/huggingface/peft}}
}

@inproceedings{wolf-etal-2020-transformers,
  title = "Transformers: State-of-the-Art Natural Language Processing",
  author = "Thomas Wolf and others",
  booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
  year = "2020"
}