LFM2.5-2.6B-JP

A Japanese fine-tuned version of LiquidAI/LFM2.5-2.6B, trained using QLoRA.

Model Details

Property	Value
Base Model	LiquidAI/LFM2.5-2.6B
Model Type	LFM2 (Liquid Foundation Model 2.5) — Dense, not MoE
Architecture	`Lfm2ForCausalLM`
Parameters	~2.6B
Hidden Size	2048
Layers	30 (mixed convolution + full attention)
Attention Heads	32 (8 KV heads, GQA)
Context Length	131,072 tokens
Vocab Size	128,000
Precision	BF16 (merged), Q4_K_M / Q8_0 / BF16 (GGUF)
License	CC-BY-NC-SA-4.0 (inherited from training data)

Training Configuration

Key Settings

Parameter	Value
Method	QLoRA (4-bit) + Unsloth
LoRA Rank (r)	16
LoRA Alpha	32
LoRA Dropout	0
Target Modules	Auto-detected: `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`, `in_proj`, `out_proj`, `w1`, `w2`, `w3`
Sequence Length	4096
Epochs	3
Learning Rate	2e-4
Gradient Accumulation	8 (effective batch size = 8)
Optimizer	`paged_adamw_8bit`
LR Scheduler	Cosine with 50 warmup steps
Weight Decay	0.01
Mixed Precision	BF16 (RTX 3060 supports it)
Gradient Checkpointing	Unsloth

Chat Template

Uses LFM2.5 native ChatML-like format (not qwen3-instruct):

<|startoftext|>|user|
User message
|assistant|
Assistant response
|user|
Next user message
|assistant|
...

Special tokens:

<|startoftext|> (BOS, id 124894)
|user| / |assistant| — role delimiters (in tokenizer as newline + special tokens)
<|pad|> (id 124893), <|endoftext|> (id 124895)

The training script verified the native template at runtime (preflight check) and uses train_on_responses_only with instruction_part="|user|" and response_part="|assistant|".

Dataset

Only msfm/ichikara-instruction-all (CC-BY-NC-SA-4.0).

Format: text (instruction) + output (response) → converted to chat template.
No data mixing; pure ichikara for license clarity (CC-BY-NC-SA-4.0).

Usage

Hugging Face Transformers (Python)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "/path/to/LFM2.5-2.6B-JP-0811"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,  # Required for Lfm2ForCausalLM
)

messages = [
    {"role": "user", "content": "日本語で自己紹介してください。"}
]

# Apply chat template
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

llama.cpp (GGUF)

# Recommended: Q4_K_M (best quality/size ratio)
./llama-cli -m LFM2.5-2.6B-JP-Q4_K_M.gguf -p "<|startoftext|>|user|
日本語で自己紹介してください。
|assistant|
" -n 512 -cnv

Ollama (via Modelfile)

FROM ./LFM2.5-2.6B-JP-Q4_K_M.gguf
TEMPLATE """{{- bos_token -}}
{{- if .System }}|system|
{{ .System }}
|assistant|
{{- end }}
{{- range .Messages }}
{{- if eq .Role "user" }}|user|
{{ .Content }}
|assistant|
{{- else }}{{ .Content }}
|assistant|
{{- end }}
{{- end }}"""
PARAMETER stop "|user|"
PARAMETER stop "|assistant|"
PARAMETER stop "

"