Note: this repo contains the LoRA adapter only (adapter_model.safetensors + adapter_config.json), not a full standalone model. Load it on top of the base model with peft, or merge it once and use it as a regular causal LM (see below).

Install

pip install -U torch transformers peft accelerate

Quick start — load the adapter on top of the base model

python

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

base_id    = "unsloth/LFM2.5-1.2B-Instruct"
adapter_id = "MenemAI/lfm-finetuned"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="cuda",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

generator = pipeline("text-generation", model=model, tokenizer=tokenizer)

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
output = generator(
    [{"role": "user", "content": question}],
    max_new_tokens=512,
    return_full_text=False,
)[0]
print(output["generated_text"])

CPU-only? Drop device_map="cuda" and pass device_map="cpu" (or "auto"); generation will be slow but works.

Run on Hugging Face Jobs

The script below works as-is with hf jobs uv run. The PEP 723 header makes uv install the right deps inside the job.

python

# /// script
# requires-python = ">=3.10"
# dependencies = [
#   "torch",
#   "transformers",
#   "peft",
#   "accelerate",
# ]
# ///
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from peft import PeftModel

base_id    = "unsloth/LFM2.5-1.2B-Instruct"
adapter_id = "MenemAI/lfm-finetuned"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype="auto", device_map="cuda", trust_remote_code=True
)
model = PeftModel.from_pretrained(base, adapter_id).eval()

generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
print(generator(
    [{"role": "user", "content": "Hello!"}],
    max_new_tokens=512,
    return_full_text=False,
)[0]["generated_text"])

hf jobs uv run --flavor a10g-small ./test.py

Optional — merge the adapter into the base model

If you want a single self-contained checkpoint (faster cold start, no peft at inference time):

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "unsloth/LFM2.5-1.2B-Instruct", torch_dtype="auto", trust_remote_code=True
)
merged = PeftModel.from_pretrained(base, "MenemAI/lfm-finetuned").merge_and_unload()
merged.save_pretrained("lfm-merged")
AutoTokenizer.from_pretrained("MenemAI/lfm-finetuned", trust_remote_code=True).save_pretrained("lfm-merged")

After merging you can load it with a plain pipeline("text-generation", model="./lfm-merged", device="cuda") or push it to a new repo with hf upload <your-user>/lfm-merged ./lfm-merged.

Training

Method: SFT via TRL
Base model: unsloth/LFM2.5-1.2B-Instruct
Dataset: mlabonne/FineTome-100k
Acceleration: Unsloth
Infrastructure: Hugging Face Jobs

Framework versions

TRL: 0.22.2
Transformers: 4.57.3
PyTorch: 2.10.0
Datasets: 4.3.0
Tokenizers: 0.22.2
PEFT: required at inference time when loading the adapter directly

Citations

bibtex

@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}}
}