| --- |
| license: other |
| license_name: lfm-open-license-v1.0 |
| base_model: |
| - LiquidAI/LFM2.5-350M-Base |
| datasets: |
| - BananaMind/BananaMind-Chat-Titles-200K |
| language: |
| - en |
| library_name: transformers |
| pipeline_tag: text-generation |
| tags: |
| - chat-title-generation |
| - title-generation |
| - conversation-title |
| - lfm |
| - lfm2.5 |
| - liquid-ai |
| - full-finetune |
| - bananamind |
| model-index: |
| - name: BananaMind Title 1.0 |
| results: [] |
| --- |
| |
|
|
|  |
| # BananaMind Title 1.0 |
|
|
| BananaMind Title 1.0 is a compact chat title generation model full-finetuned from LiquidAI/LFM2.5-350M-Base. |
|
|
| ## Examples |
|
|
| Message: |
|
|
| why is my wifi disconnecting |
| |
| Title: |
|
|
| Troubleshooting Wi-Fi Disconnection Issues |
| |
| Message: |
|
|
| write a c# method to balance a budget |
| |
| Title: |
|
|
| C# Budget Balancing Method |
| |
|
|
| The model is intended to create neutral conversation titles. It is not intended to answer the user request. |
|
|
| ## Model Details |
|
|
| - Model name: BananaMind Title 1.0 |
| - Base model: LiquidAI/LFM2.5-350M-Base |
| - Model family: LFM2.5 |
| - Base parameter count: 350M |
| - Training method: full finetune |
| - Precision: BF16 |
| - Task: first user message to short chat title |
| - Dataset: BananaMind/BananaMind-Chat-Titles-200K |
| - Recommended max input length: 2048 tokens |
|
|
|
|
| <h2 align="center">Why BananaMind Title?</h2> |
| Most AI Platforms use their most powerful model to generate titles, While this works, it uses alot of compute just for the title. |
| BananaMind Title Focuses only on title generation and makes it only 350M Parameters which dramatically decreases the compute needed, this thing runs on a CPU Only Phone! |
|
|
| ## GGUF is coming soon. |
|
|
|
|
| ## Prompt Format |
|
|
| Use this prompt format for best results: |
|
|
| Generate a short chat title. |
| Message: <first user message> |
| Title: |
| |
| The model should complete only the title. |
|
|
| <h3 align="center">Using BananaMind Title</h3> |
|
|
| import torch |
| from transformers import AutoTokenizer, AutoModelForCausalLM |
| |
| model_id = "BananaMind/BananaMind-Title-1.0" |
| |
| tokenizer = AutoTokenizer.from_pretrained( |
| model_id, |
| trust_remote_code=True, |
| ) |
| |
| if tokenizer.pad_token is None: |
| tokenizer.pad_token = tokenizer.eos_token |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| dtype=torch.bfloat16, |
| device_map="auto", |
| trust_remote_code=True, |
| ) |
| |
| model.config.use_cache = True |
| model.eval() |
| |
| message = "What is a AI?" |
| |
| prompt = f"""Generate a short chat title. |
| Message: {message} |
| Title:""" |
| |
| inputs = tokenizer( |
| prompt, |
| return_tensors="pt", |
| truncation=True, |
| max_length=768, |
| ).to(model.device) |
| |
| with torch.inference_mode(): |
| output = model.generate( |
| **inputs, |
| max_new_tokens=12, |
| do_sample=False, |
| repetition_penalty=1.1, |
| eos_token_id=tokenizer.eos_token_id, |
| pad_token_id=tokenizer.pad_token_id, |
| use_cache=True, |
| ) |
| |
| title = tokenizer.decode( |
| output[0][inputs["input_ids"].shape[1]:], |
| skip_special_tokens=True, |
| ).strip().split("\n")[0].strip() |
| |
| print(title) |
| |
| ## Training Summary |
|
|
| BananaMind Title 1.0 was full-finetuned from LiquidAI/LFM2.5-350M-Base on BananaMind/BananaMind-Chat-Titles-200K. |
|
|
| Training format: |
|
|
| Generate a short chat title. |
| Message: <first user message> |
| Title: <target title> |
| |
| Prompt/input tokens were masked from loss. The model was trained to predict only the title completion tokens. |
|
|
| Training setup: |
|
|
| - base model: LiquidAI/LFM2.5-350M-Base |
| - dataset: BananaMind/BananaMind-Chat-Titles-200K |
| - method: full finetune |
| - precision: BF16 |
| - max sequence length: 2048 |
| - learning rate: 3e-5 |
| - effective batch size: 128 |
| - stopped around epoch 1 |
| - observed eval loss near selected checkpoint: about 0.769 |
|
|
|
|
|
|
| ## License and Attribution |
|
|
| This model is a derivative fine-tune of LiquidAI/LFM2.5-350M-Base by Liquid AI. |
|
|
| The base model is released under the LFM Open License v1.0. This fine-tuned model is distributed under the same license family and remains subject to the LFM Open License v1.0 terms. |
|
|
| Attribution: |
|
|
| - Base model: LiquidAI/LFM2.5-350M-Base |
| - Base model creator: Liquid AI, Inc. |
| - Base model license: LFM Open License v1.0 |
| - Fine-tuned model: BananaMind/BananaMind-Title-1.0 |
| - Fine-tuned by: BananaMind |
| - Training dataset: BananaMind/BananaMind-Chat-Titles-200K |
|
|
| Liquid AI retains ownership of the original base model. BananaMind provides the fine-tuning modifications for this derivative model. |
|
|
| Users should review the LFM Open License v1.0 before using, redistributing, or commercially deploying this model. |
|
|
| ## Citation |
|
|
| If you use this model, please attribute both the base model and this fine-tune: |
|
|
| BananaMind Title 1.0 is a full fine-tune of LiquidAI/LFM2.5-350M-Base for chat title generation, trained on BananaMind/BananaMind-Chat-Titles-200K. |
| |
| Please retain Liquid AI attribution and license notices for the LFM base model. |
|
|
