Instructions to use KeisukeMiyamoto/lfm2.5-230m-cpt with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use KeisukeMiyamoto/lfm2.5-230m-cpt with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="KeisukeMiyamoto/lfm2.5-230m-cpt")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("KeisukeMiyamoto/lfm2.5-230m-cpt")
model = AutoModelForCausalLM.from_pretrained("KeisukeMiyamoto/lfm2.5-230m-cpt", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use KeisukeMiyamoto/lfm2.5-230m-cpt with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "KeisukeMiyamoto/lfm2.5-230m-cpt"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KeisukeMiyamoto/lfm2.5-230m-cpt",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/KeisukeMiyamoto/lfm2.5-230m-cpt

SGLang

How to use KeisukeMiyamoto/lfm2.5-230m-cpt with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "KeisukeMiyamoto/lfm2.5-230m-cpt" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KeisukeMiyamoto/lfm2.5-230m-cpt",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "KeisukeMiyamoto/lfm2.5-230m-cpt" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KeisukeMiyamoto/lfm2.5-230m-cpt",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use KeisukeMiyamoto/lfm2.5-230m-cpt with Docker Model Runner:
```
docker model run hf.co/KeisukeMiyamoto/lfm2.5-230m-cpt
```

LFM2.5-230M-CPT

LFM2.5-230M-CPT is a Japanese text continuation model obtained by continued pretraining of LiquidAI/LFM2.5-230M-Base on speech transcripts. It models the wording and style of Japanese spoken language. Despite the transcript source, this checkpoint has no audio input or audio encoder and cannot transcribe speech by itself. It is a base model for further adaptation, not an instruction-tuned assistant.

Model Details

Item	Value
Architecture	`Lfm2ForCausalLM`, decoder-only LFM2 hybrid
Parameters	229,693,184 serialized parameters (about 230M)
Layers	14: 8 causal-convolution blocks and 6 grouped-query attention blocks
Hidden size	1,024
Feed-forward intermediate size	2,560
Attention heads / key-value heads	16 / 8
Model vocabulary capacity	65,536 token IDs
Tokenizer	Original LFM2.5-230M-Base tokenizer, unchanged
Base-model advertised context length	32,768 tokens
Maximum sequence length used in continued pretraining	512 tokens, including the end-of-sequence token
Weight format	BF16 Safetensors (`model.safetensors`, about 459 MB)
Framework	Hugging Face Transformers

The configuration retains the base model's positional-embedding setting, but this continued-pretraining run used sequences of at most 512 tokens. Performance on longer contexts has not been evaluated for this checkpoint.

Training Data

The continued-pretraining corpus is the text field of KeisukeMiyamoto/audio-transcript, a Japanese text-only dataset assembled from NHK radio news, NHK Archives, and technology podcasts. The source transcripts include machine-generated text and corrections. No audio, prev_text, token-count columns, or other dataset fields were supplied to the model.

Split	Role	Transcript rows	Dataset `text_token` total
`train`	Continued pretraining	1,143,151	70,156,912
`validation`	Checkpoint selection	63,260	3,897,592
`test`	Not used in training or checkpoint selection	63,253	3,897,621

The dataset token totals exclude special tokens. Each transcript was tokenized with the base-model tokenizer and followed by one end-of-sequence token for training. Transcripts were packed into training sequences. Dataset splits are separated by row, not by recording, episode, or speaker; related segments may appear in different splits. See the dataset card for construction details and source-rights information.

Training Configuration

Item	Value
Method	Full-parameter causal language-model continued pretraining
Objective	Next-token prediction on transcript text
Epochs completed	3
Optimizer	Fused AdamW
Peak learning rate	5 × 10⁻⁵
Learning-rate schedule	Cosine decay with 10% warmup
Per-device batch size	72 configured; automatic batch-size adjustment enabled
Gradient accumulation	2 steps
Precision during training	BF16 mixed precision
Gradient checkpointing	Enabled
Validation frequency	Once per epoch
Optimizer steps completed	23,817
Published checkpoint	Best validation checkpoint at step 15,878 (end of epoch 2)

The selected checkpoint was saved in BF16 for distribution. The published configuration enables generation caching (use_cache=true). The tokenizer and model architecture remain compatible with standard Transformers from_pretrained() loading.

Evaluation

Validation loss is token-weighted next-token cross-entropy on the validation split. Lower is better.

Epoch	Step	Validation loss
1	7,939	3.0819
2	15,878	2.9808
3	23,817	2.9848

The epoch-2 checkpoint was selected because it had the lowest validation loss. This is a text language-modeling measurement, not an ASR metric. No audio-to-text evaluation, word error rate, character error rate, or standardized language-model benchmark is reported for this checkpoint.

Usage

Tested with Transformers 5.17.0, PyTorch 2.14.0, and python-dotenv 1.2.3. Set HF_TOKEN in a local .env file. The example performs greedy continuation on a CUDA device:

import torch
from dotenv import dotenv_values
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "KeisukeMiyamoto/lfm2.5-230m-cpt"
token = dotenv_values(".env")["HF_TOKEN"]
tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, token=token).to("cuda")
inputs = tokenizer("今日は天気が良いので、", return_tensors="pt").to("cuda")

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=48, do_sample=False)

continuation = tokenizer.decode(
    output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
)
print(continuation)

This model repository is private at the time of writing. Authorized users must authenticate with Hugging Face before loading it. The included chat template is inherited from the base model; the continued-pretraining data contained plain transcripts, not chat instruction/response pairs.

Example Outputs

These are illustrative greedy continuations from the published BF16 checkpoint with max_new_tokens=48; they are not benchmark results.

Prompt	Generated continuation
`今日は天気が良いので、`	`ちょっとお休みさせていただきます。`
`この問題について、私は`	`この問題を解決するにはどうしたらいいかということを考えて、`

Limitations

No audio support: an audio encoder and paired audio-text training would be required for speech recognition.
The dataset concentrates on Japanese broadcast and podcast speech. It is not representative of all Japanese domains or writing styles.
Transcripts may contain recognition errors, omissions, unusual punctuation, and content or speaker biases inherited from the sources.
Text may end abruptly, repeat phrases, or produce plausible but unsupported continuations. The end-of-sequence token can cause short or empty outputs.
The model has not been instruction-tuned or safety-aligned. Do not assume chat behavior from the inherited chat template.
Related recordings can span dataset splits, so the reported validation loss should not be interpreted as a source-independent generalization score.

License and Source Rights

The model inherits the base model's LFM 1.0 license. The transcript dataset and its source collections have separate rights and use conditions. Consult their dataset cards and underlying content owners before reusing or redistributing the training text. This model card does not grant rights to the source recordings or transcripts.