91 lines
2.9 KiB
Markdown
91 lines
2.9 KiB
Markdown
|
---
|
|||
|
|
base_model: unsloth/Qwen2.5-1.5B-Instruct
|
|||
|
|
tags:
|
|||
|
|
- text-generation-inference
|
|||
|
|
- transformers
|
|||
|
|
- unsloth
|
|||
|
|
- qwen2.5
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
- hi
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
As calling operations scale, it becomes clear that dialing and talking is not enough.
|
|||
|
|
Even with a strong voice AI + telephony architecture, the real value shows up only when post-call actions are captured and executed in a robust, dependable and consistent way. Closing the loop matters more than just connecting the call.
|
|||
|
|
|
|||
|
|
To support that, we’re releasing our Hindi + English transcript analytics model tuned specifically for call transcripts:
|
|||
|
|
|
|||
|
|
|
|||
|
|
You can plug it into your calling or voice AI stack to automatically extract:
|
|||
|
|
|
|||
|
|
• Enum-based classifications (e.g., call outcome, intent, disposition)
|
|||
|
|
• Conversation summaries
|
|||
|
|
• Action items / follow-ups
|
|||
|
|
|
|||
|
|
It’s built to handle real-world Hindi, English, and mixed Hinglish calls, including noisy transcripts.
|
|||
|
|
|
|||
|
|
Finetuning Parameters:
|
|||
|
|
```
|
|||
|
|
rank = 64
|
|||
|
|
lora_alpha = rank*2,
|
|||
|
|
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
|
|||
|
|
"gate_proj", "up_proj", "down_proj",],
|
|||
|
|
SFTConfig(
|
|||
|
|
dataset_text_field = "prompt",
|
|||
|
|
per_device_train_batch_size = 32,
|
|||
|
|
gradient_accumulation_steps = 1, # Use GA to mimic batch size!
|
|||
|
|
warmup_steps = 5,
|
|||
|
|
num_train_epochs = 2,
|
|||
|
|
learning_rate = 2e-4,
|
|||
|
|
logging_steps = 50,
|
|||
|
|
optim = "adamw_8bit",
|
|||
|
|
weight_decay = 0.001,
|
|||
|
|
lr_scheduler_type = "linear",
|
|||
|
|
seed = SEED,
|
|||
|
|
report_to = "wandb",
|
|||
|
|
eval_strategy="steps",
|
|||
|
|
eval_steps=200,
|
|||
|
|
)
|
|||
|
|
The model was finetuned on ~100,000 curated transcripts across different domanins and language preferences
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
Provide the below schema for best output:
|
|||
|
|
```
|
|||
|
|
response_schema = {
|
|||
|
|
"type": "object",
|
|||
|
|
"properties": {
|
|||
|
|
"key_points": {
|
|||
|
|
"type": "array",
|
|||
|
|
"items": {"type": "string"},
|
|||
|
|
"nullable": True,
|
|||
|
|
},
|
|||
|
|
"action_items": {
|
|||
|
|
"type": "array",
|
|||
|
|
"items": {"type": "string"},
|
|||
|
|
"nullable": True,
|
|||
|
|
},
|
|||
|
|
"summary": {"type": "string"},
|
|||
|
|
"classification": classification_schema,
|
|||
|
|
},
|
|||
|
|
"required": ["summary", "classification"],
|
|||
|
|
}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
- **Developed by:** RinggAI
|
|||
|
|
- **License:** apache-2.0
|
|||
|
|
- **Finetuned from model :** unsloth/Qwen2.5-1.5B-Instruct
|
|||
|
|
- Parameter decision where made using
|
|||
|
|
**Schulman, J., & Thinking Machines Lab. (2025).**
|
|||
|
|
*LoRA Without Regret.*
|
|||
|
|
Thinking Machines Lab: Connectionism.
|
|||
|
|
DOI: 10.64434/tml.20250929
|
|||
|
|
Link: https://thinkingmachines.ai/blog/lora/
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
[<img style="border-radius: 20px;" src="https://storage.googleapis.com/desivocal-prod/desi-vocal/logo.png" width="200"/>](https://ringg.ai)
|
|||
|
|
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|