A supervised fine-tune of unsloth/LFM2.5-230M on the kth8/title-generation-10000x dataset. Supports all 10 languages by the base model: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish.
Trained with the system prompt used by OpenCode's title agent. OpenCode generate thread title based on the first user message. OpenCode UI can show title up to 100 characters long before being truncated.
Usage example
Point to this model with small_model key in opencode.jsonc config file.
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"title": {
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "http://127.0.0.1:8080/v1",
"apiKey": "not-needed"
},
"models": {
"generator": {}
}
}
},
"small_model": "title/generator"
}
User prompt
Troy's home is 75 meters away from school while Emily's home is 98 meters away from school. Troy and Emily walk to school and back home every day. How much farther does Emily walk to school and back in five days?
Title response
Troy vs Emily five-day school walk distance difference
Training Settings
PEFT
- Rank: 16
- LoRA alpha: 16
- Modules: ["q_proj", "k_proj", "v_proj", "out_proj", "in_proj", "w1", "w2", "w3"]
- Gradient checkpointing: unsloth
SFT
- Epoch: 1
- Batch size: 4
- Gradient Accumulation steps: 4
- Learning rate: 0.0002
- Optimizer: adamw_torch_fused
- Learning rate scheduler: cosine
- Warmup steps: 10
- Weight decay: 0.01
Framework versions
- Unsloth: 2026.9.4
- TRL: 0.23.1
- Transformers: 5.5.0
- Pytorch: 2.10.0+cu128
- Datasets: 4.3.0
- Tokenizers: 0.22.2
License
This model is released under LFM Open License v1.0.