File size: 2,825 Bytes
1186208
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
---
language: en
license: apache-2.0
library_name: transformers
tags:
  - tptt
  - peft
  - trust_remote_code
pipeline_tag: text-generation
base_model: meta-llama/Llama-3.2-1B
datasets:
- yahma/alpaca-cleaned
---

# lora_delta_product_r_m0.5_constant

<p align="center">
    <a href="https://arxiv.org/abs/2506.17671">
        <img alt="arXiv" src="https://img.shields.io/badge/arXiv-tptt-blueviolet.svg">
    </a>
    <a href="https://pypi.org/project/tptt/">
        <img alt="PyPI" src="https://img.shields.io/pypi/v/tptt?color=orange">
    </a>
    <a href="https://github.com/fabienfrfr/tptt/">
        <img alt="Release" src="https://img.shields.io/github/v/release/fabienfrfr/tptt?color=brightgreen">
    </a>
    <a href="https://fabienfrfr.github.io/tptt/">
        <img alt="Documentation" src="https://img.shields.io/badge/docs-online-blue">
    </a>
    <a href="https://huggingface.co/ffurfaro">
        <img alt="HuggingFace" src="https://img.shields.io/badge/hf-ffurfaro-yellow">
    </a>
</p>

Titanesque version of `meta-llama/Llama-3.2-1B` with parallel linearized attention (TPTT 😊) and PEFT.

The architecture was presented in the paper [TPTT](https://huggingface.co/papers/2506.17671).


## Model Details

- **Architecture:** ['TpttModel']
- **Base model:** meta-llama/Llama-3.2-1B
- **LiZA config:** operator=delta_product_r, mag=0.5
- **LoRA config:** r=8, alpha=16, dropout=0.05
- **torch_dtype:** 

## Usage


```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
"ffurfaro/lora_delta_product_r_m0.5_constant",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("ffurfaro/meta-llama/Llama-3.2-1B")

prompt = "Your prompt here"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs, skip_special_tokens=True))

```

> [!IMPORTANT]
> You must specify the `subfolder` if the repo contains multiple models, see the homepage for details.

## Training

- **Dataset:** yahma/alpaca-cleaned
- **Platform:** Kaggle
- **Hardware:** 2xT4
- **Batch size:** 2
- **Epochs:** 1.0
- **Learning rate (final):** N/A
- **Loss (final):** 7.606347968441995
- **Training runtime:** 2004.1174 sec
- **Samples per second:** 1.291
- **Steps per second:** 0.323
- **Total FLOPs:** 1937596357804032.0
- **Gradient norm (final):** N/A

## Evaluation

- **Metrics:** Training loss only (no eval yet, table soon : PiQA, ARC, Hella, Wino, GSM8K, MMLU)
- **Results:** Final training loss: 7.606347968441995


## Citation & Contact

If you use TPTT in your academic work, please cite [Furfaro](https://huggingface.co/ffurfaro). For questions or support, please open an issue on the [GitHub repository](https://github.com/fabienfrfr/tptt) or contact the maintainer.


---