OpenIntelligenceNet's picture
Add comprehensive Model Card and metadata tags
8e7cf5c verified
|
Raw History Blame Contribute Delete
3.35 kB
---
base_model: LiquidAI/LFM3-1.7B
library_name: transformers
pipeline_tag: text-generation
tags:
- liquid-ai
- lfm3
- heretic
- abliterated
- uncensored
- decensored
- text-generation
- conversational
- causal-lm
- hybrid-architecture
- edge-ai
- roleplay
- reasoning
- custom_code
---
# LFM3-1.7B-Uncensored
**`OpenIntelligenceNet/LFM3-1.7B-Uncensored`** is an uncensored, directional-abliterated variant of **[LiquidAI/LFM3-1.7B](https://huggingface.co/LiquidAI/LFM3-1.7B)** created using **[Heretic](https://github.com/p-e-w/heretic)** (`v2.0.0.dev0`).
By leveraging multi-objective Pareto optimization via Optuna (100 trials), this model removes synthetic refusal directions across the residual stream while preserving the base model's instruction-following, reasoning, and low KL-divergence characteristics.
---
## Key Highlights
* **Base Architecture:** Liquid Foundation Model 3 (`LFM3-1.7B`) featuring a fast, memory-efficient tri-hybrid architecture (gated short convolutions, sliding-window attention, and global attention layers).
* **Refusal Abliteration:** Processed with Heretic's parameterized directional abliteration (`heretic.modifiers.abliteration.Abliteration`) to eliminate pre-programmed moralizing and unnecessary refusals.
* **Out-of-the-Box Compatibility:** Includes a patched `modeling_lfm3.py` so `trust_remote_code=True` works seamlessly across standard `transformers` versions without `linear_attention` mask errors.
* **Lightweight & Fast:** Ideal for local deployment, creative writing, unfiltered roleplay, red-teaming research, and edge inference.
---
## Quickstart (Transformers)
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "OpenIntelligenceNet/LFM3-1.7B-Uncensored"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float16,
device_map="auto",
trust_remote_code=True
)
messages = [
{"role": "user", "content": "Explain the difference between a process and a thread in operating systems."}
]
input_text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.2,
top_k=50,
repetition_penalty=1.1,
do_sample=True,
pad_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```
## Abliteration Details
* **Tooling:** Heretic v2.0.0.dev0
* **Modifier:** heretic.modifiers.abliteration.Abliteration
* **Optimization:** 100 Optuna trials (30 startup trials, batch_size=8, max_response_length=100)
* **Export Strategy:** Full weight merge (MERGE)
## Attribution & Usage
This model is a derivative work of LiquidAI/LFM3-1.7B and remains subject to the original LFM Open License v1.0. Because safety guardrails and refusal mechanisms have been removed, users are solely responsible for the outputs generated and for ensuring safe, legal deployment in downstream applications.