File size: 1,353 Bytes
1ce2b6a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language:
- es
---
## EQASpa 7b

This model is a fine-tuned version of [Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) for the Extractive Question Answering task in spanish.

It was fine-tuned on the [QuALES 2022](https://www.fing.edu.uy/inco/grupos/pln/quales/) dataset training partition using LoRA for one epoch.

## Prompt format

To use the model, the following prompting format should be applied:


    ### TEXTO:
    {{Context document}}
    ### PREGUNTA:
    {{Question}}
    ### RESPUESTA:


## Evaluation

We evaluate the model on the test partition of the QuALES dataset, and compare it with one-shot prompting as a baseline.

 Prompt | Model | Acc_exact | F_bertscore
 --- | --- | --- | ---
one-shot prompting | [zephyr-7b-beta](HuggingFaceH4/zephyr-7b-beta) | 0.025 | 0.614
one-shot prompting | [Llama-2-13b-chat-hf](meta-llama/Llama-2-13b-chat-hf) | 0.192 | 0.700
default | EQASpa 7b | **0.225** | **0.713**


## Training procedure


The following `bitsandbytes` quantization config was used during training:
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: False
- bnb_4bit_compute_dtype: float16

### Framework versions


- PEFT 0.4.0