LFM2.5 (1.2B) Translation
To run this, press "Runtime" and press "Run all" on a free Tesla T4 Google Colab instance!
To install Unsloth on your local device, follow our guide. This notebook is licensed LGPL-3.0.
You will learn how to do data prep, how to train, how to run the model, & how to save it
News
Train MoEs - DeepSeek, GLM, Qwen and gpt-oss 12x faster with 35% less VRAM. Blog
You can now train embedding models 1.8-3.3x faster with 20% less VRAM. Blog
Ultra Long-Context Reinforcement Learning is here with 7x more context windows! Blog
3x faster LLM training with 30% less VRAM and 500K context. 3x faster • 500K Context
New in Reinforcement Learning: FP8 RL • Vision RL • Standby • gpt-oss RL
Visit our docs for all our model uploads and notebooks.
Installation
Unsloth
env: UNSLOTH_RETURN_LOGITS=1 # Run this to disable CCE since it is not supported for CPT
🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning. 🦥 Unsloth Zoo will now patch everything to make training faster! ==((====))== Unsloth 2025.12.8: Fast Lfm2 patching. Transformers: 4.57.3. \\ /| NVIDIA L4. Num GPUs = 1. Max memory: 22.161 GB. Platform: Linux. O^O/ \_/ \ Torch: 2.9.0+cu126. CUDA: 8.9. CUDA Toolkit: 12.6. Triton: 3.5.0 \ / Bfloat16 = TRUE. FA [Xformers = 0.0.33.post1. FA2 = False] "-____-" Free license: http://github.com/unslothai/unsloth Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored! Unsloth: QLoRA and full finetuning all not selected. Switching to 16bit LoRA.
model.safetensors: 0%| | 0.00/2.34G [00:00<?, ?B/s]
generation_config.json: 0%| | 0.00/132 [00:00<?, ?B/s]
tokenizer_config.json: 0.00B [00:00, ?B/s]
tokenizer.json: 0.00B [00:00, ?B/s]
special_tokens_map.json: 0%| | 0.00/434 [00:00<?, ?B/s]
chat_template.jinja: 0.00B [00:00, ?B/s]
We now add LoRA adapters so we only need to update 1 to 10% of all parameters!
We also add embed_tokens and lm_head to allow the model to learn out of distribution data.
/usr/local/lib/python3.12/dist-packages/peft/tuners/tuners_utils.py:916: UserWarning: Model with `tie_word_embeddings=True` and the tied_target_modules=['model.embed_tokens', 'lm_head'] are part of the adapter. This can lead to complications, for example when merging the adapter or converting your model to formats other than safetensors. See for example https://github.com/huggingface/peft/issues/2018. warnings.warn(
Unsloth: Making `model.base_model.model.model.embed_tokens` require gradients
Data Prep
We now use the Korean subset of the Wikipedia dataset to first continually pretrain the model. You can use any language you like! Go to Wikipedia's List of Languages to find your own language!
[NOTE] To train only on completions (ignoring the user's input) read our docs here
[NOTE] Remember to add the EOS_TOKEN to the tokenized output! Otherwise you'll get infinite generations!
If you want to use the llama-3 template for ShareGPT datasets, try our conversational notebook
For text completions like novel writing, try this notebook.
[NOTE] Use https://translate.google.com to translate from English to Korean!
We only use 1% of the dataset to speed things up! Use more for longer runs!
README.md: 0.00B [00:00, ?B/s]
20231101.ko/train-00000-of-00003.parquet: 0%| | 0.00/400M [00:00<?, ?B/s]
20231101.ko/train-00001-of-00003.parquet: 0%| | 0.00/205M [00:00<?, ?B/s]
20231101.ko/train-00002-of-00003.parquet: 0%| | 0.00/177M [00:00<?, ?B/s]
Generating train split: 0%| | 0/647897 [00:00<?, ? examples/s]
Map: 0%| | 0/6478 [00:00<?, ? examples/s]
Continued Pretraining
Now let's use Unsloth's UnslothTrainer! More docs here: TRL SFT docs. We do 20 steps to speed things up, but you can set num_train_epochs=1 for a full run, and turn off max_steps=None.
Also set embedding_learning_rate to be a learning rate at least 2x or 10x smaller than learning_rate to make continual pretraining work!
Unsloth: Tokenizing ["text"] (num_proc=16): 0%| | 0/6478 [00:00<?, ? examples/s]
🦥 Unsloth: Padding-free auto-enabled, enabling faster training.
GPU = NVIDIA L4. Max memory = 22.161 GB. 2.598 GB of memory reserved.
The model is already on multiple devices. Skipping the move to device specified in `args`. ==((====))== Unsloth - 2x faster free finetuning | Num GPUs used = 1 \\ /| Num examples = 6,478 | Num Epochs = 1 | Total steps = 120 O^O/ \_/ \ Batch size per device = 2 | Gradient accumulation steps = 8 \ / Data Parallel GPUs = 1 | Total batch size (2 x 8 x 1) = 16 "-____-" Trainable parameters = 97,517,568 of 1,276,508,928 (7.64% trained)
Unsloth: Will smartly offload gradients to save VRAM!
/usr/local/lib/python3.12/dist-packages/peft/utils/save_and_load.py:279: UserWarning: Setting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.
warnings.warn("Setting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.") Instruction Finetuning
We now use the Alpaca in GPT4 Dataset but translated in Korean!
Go to vicgalle/alpaca-gpt4 for the original GPT4 dataset for Alpaca or MultilingualSIFT project for other translations of the Alpaca dataset.
README.md: 0%| | 0.00/124 [00:00<?, ?B/s]
Repo card metadata block was not found. Setting CardData to empty. WARNING:huggingface_hub.repocard:Repo card metadata block was not found. Setting CardData to empty.
alpaca-gpt4-korean.json: 0%| | 0.00/51.6M [00:00<?, ?B/s]
Generating train split: 0%| | 0/49969 [00:00<?, ? examples/s]
We print 1 example:
{'conversations': [{'from': 'human', 'value': '재활용 캠페인 슬로건을 제시하세요.\n'}, {'from': 'gpt', 'value': '1. "더욱 녹색 미래를 위해 함께 줄이고, 재사용하고, 재활용하세요."\n2. "더 나은 내일을 위해 오늘 바로 재활용하세요."\n3. "쓰레기를 보물로 만드는 법 - 재활용!"\n4. "인생의 순환을 위해 재활용하세요."\n5. "자원을 아끼고 더 많이 재활용하세요."'}], 'id': '23712'}
We again use https://translate.google.com/ to translate the Alpaca format into Korean
Map: 0%| | 0/49969 [00:00<?, ? examples/s]
We again employ UnslothTrainer and do instruction finetuning!
Unsloth: Tokenizing ["text"] (num_proc=16): 0%| | 0/49969 [00:00<?, ? examples/s]
The model is already on multiple devices. Skipping the move to device specified in `args`. ==((====))== Unsloth - 2x faster free finetuning | Num GPUs used = 1 \\ /| Num examples = 49,969 | Num Epochs = 1 | Total steps = 120 O^O/ \_/ \ Batch size per device = 2 | Gradient accumulation steps = 8 \ / Data Parallel GPUs = 1 | Total batch size (2 x 8 x 1) = 16 "-____-" Trainable parameters = 97,517,568 of 1,276,508,928 (7.64% trained)
Unsloth: Will smartly offload gradients to save VRAM!
/usr/local/lib/python3.12/dist-packages/peft/utils/save_and_load.py:279: UserWarning: Setting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.
warnings.warn("Setting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.") 226.8895 seconds used for training. 3.78 minutes used for training. Peak reserved memory = 4.361 GB. Peak reserved memory for training = 1.763 GB. Peak reserved memory % of max memory = 19.679 %. Peak reserved memory for training % of max memory = 7.955 %.
Inference
Let's run the model! You can change the instruction and input - leave the output blank!
Remember to use https://translate.google.com/!
['<|startoftext|>다음은 작업을 설명하는 명령입니다. 요청을 적절하게 완료하는 응답을 작성하세요.\n\n### 지침:\n피보나치 수열을 계속하세요: 1, 1, 2, 3, 5, 8,\n\n### 응답:\n피보나치 수열은 각 숫자가 앞의 두 숫자의 합인 수열입니다. 이 수열은 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, ']
You can also use a TextStreamer for continuous inference - so you can see the generation token by token, instead of waiting the whole time!
<|startoftext|>다음은 작업을 설명하는 명령입니다. 요청을 적절하게 완료하는 응답을 작성하세요. ### 지침: 한국음악은 어떤가요? ### 응답: 한국음악은 전통 음악과 현대 음악을 포함한 다양한 장르를 포함합니다. 전통 음악은 종종 악기와 노래를 사용하여 한국의 역사와 문화를 반영합니다. 현대 음악은 전통 음악의 영향을 받아 발전해왔으며, 전자 음악, 록, 팝 등 다양한 장르가 있습니다. 한국 음악은 전 세계적으로 인기를 얻고 있으며, 많은 아티스트들이 국제적인 성공을 거두었습니다.<|im_end|>
By using https://translate.google.com/ we get
Korean music is classified into many types of music genres.
This genre is classified into different music genres such as pop songs,
rock songs, classical songs and pop songs, music groups consisting of drums, fans, instruments and singers
/usr/local/lib/python3.12/dist-packages/peft/utils/save_and_load.py:279: UserWarning: Setting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.
warnings.warn("Setting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.")('lora_model/tokenizer_config.json',
, 'lora_model/special_tokens_map.json',
, 'lora_model/chat_template.jinja',
, 'lora_model/tokenizer.json') Now if you want to load the LoRA adapters we just saved for inference, set False to True:
<|startoftext|>다음은 작업을 설명하는 명령입니다. 요청을 적절하게 완료하는 응답을 작성하세요. ### 지침: 지구를 광범위하게 설명하세요. ### 응답: 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명하세요. 지구를 광범위하게 설명
By using https://translate.google.com/ we get
Earth refers to all things including natural disasters such as local derailment
and local depletion that occur in one space along with the suppression of water, gases, and living things.
Most of the Earth's water comes from oceans, atmospheric water, underground water layers, and rivers and rivers.
Yikes the language model is a bit whacky! Change the temperature and using sampling will definitely make the output much better!
You can also use Hugging Face's AutoPeftModelForCausalLM. Only use this if you do not have unsloth installed. It can be hopelessly slow, since 4bit model downloading is not supported, and Unsloth's inference is 2x faster.
Saving to float16 for vLLM
We also support saving to float16 directly. Select merged_16bit for float16 or merged_4bit for int4. We also allow lora adapters as a fallback. Use push_to_hub_merged to upload to your Hugging Face account! You can go to https://huggingface.co/settings/tokens for your personal tokens. See our docs for more deployment options.
GGUF / llama.cpp Conversion
To save to GGUF / llama.cpp, we support it natively now! We clone llama.cpp and we default save it to q8_0. We allow all methods like q4_k_m. Use save_pretrained_gguf for local saving and push_to_hub_gguf for uploading to HF.
Some supported quant methods (full list on our docs page):
q8_0- Fast conversion. High resource use, but generally acceptable.q4_k_m- Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K.q5_k_m- Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K.
Now, use the lfm_finetune.Q8_0.gguf file or lfm_finetune.Q4_K_M.gguf file in llama.cpp.
And we're done! If you have any questions on Unsloth, we have a Discord channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!
Some other resources:
- Train your own reasoning model - Llama GRPO notebook Free Colab
- Saving finetunes to Ollama. Free notebook
- Llama 3.2 Vision finetuning - Radiography use case. Free Colab
- See notebooks for DPO, ORPO, Continued pretraining, conversational finetuning and more on our documentation!
