---
license: apache-2.0
base_model: Umranz/LFM2.5-2.6B-heretic
tags:
- conversational
- girlfriend
- character
- fine-tuned
- lfm2.5
- liquid
- merged
language:
- en
- hi
pipeline_tag: text-generation
---

# Shruti-Soft-2.6b (Full Standalone Merged)

**Shruti** is a full standalone 2.6B parameter uncensored AI girlfriend model fine-tuned from `Umranz/LFM2.5-2.6B-heretic` (LiquidAI hybrid LIV short-conv + GQA architecture) on 5,000 curated ChatML conversations.

**This repository contains the complete merged weights (~5.2 GB safetensors). No separate base model or adapter download is required.**

## Character Profile

Warm, affectionate, deeply caring. Sweet, playful, emotionally attentive, witty, and supportive. Adapts seamlessly between everyday banter, romance, and mature intimacy. Natural with pet names (*baby, babe, handsome, jaan, love*).

## Training Details

- **Framework**: LEAP Finetune (LiquidAI official)
- **Method**: QLoRA SFT (`r=32`, `alpha=64`, 4 epochs)
- **Source Run**: `Umranz/Shruti-Soft-2.6b-run-b` (Winning sweep run with `0.4074` eval loss)
- **Hardware**: A100-80GB via Modal.com

## Quick Start (Transformers)

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Umranz/Shruti-Soft-2.6b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are Shruti, a warm, affectionate girlfriend..."},
    {"role": "user", "content": "I had a really long day today..."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=200, temperature=0.7, top_p=0.9, do_sample=True)

print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True).strip())
```