mirror of
https://github.com/marvinscham/masterthesis-playground.git
synced 2026-03-22 00:12:42 +01:00
Add helper stuff for figures, cleanup
This commit is contained in:
35
raft/README.md
Normal file
35
raft/README.md
Normal file
@@ -0,0 +1,35 @@
|
||||
# Retrieval-Augmented Finetuning (RAFT)
|
||||
|
||||
**Ablauf**:
|
||||
|
||||
## Vorbereiten des Retrieval-Corpus
|
||||
|
||||
```bash
|
||||
python prepare_corpus.py --input_tab ../data/intermediate/selected_topics_documents.csv --out_dir out
|
||||
```
|
||||
|
||||
## Erstellen des RAFT-Datensatzes
|
||||
|
||||
```bash
|
||||
python make_raft_data.py --out_dir out --n_examples 100
|
||||
```
|
||||
|
||||
## Training der QLoRA-Adapter
|
||||
|
||||
```bash
|
||||
python train_mistral_raft.py --train_jsonl out/raft_train.jsonl --out_dir out/mistral_balitwin_lora
|
||||
```
|
||||
|
||||
## Inferenz
|
||||
|
||||
### Per Baseline Mistral 7B + PEFT-Adapter
|
||||
|
||||
```bash
|
||||
python rag_chat.py --lora_dir out/mistral_balitwin_lora
|
||||
```
|
||||
|
||||
### Pre-Merged Modell + Adapter
|
||||
|
||||
```bash
|
||||
python rag_chat_merged.py --model_dir /path/to/model_folder --out_dir out
|
||||
```
|
||||
@@ -10,9 +10,6 @@ from sentence_transformers import SentenceTransformer
|
||||
from tqdm import tqdm
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
## Usage: python make_raft_data.py --out_dir out --n_examples 5000
|
||||
|
||||
|
||||
SYSTEM_PERSONA = """You are 'BaliTwin', a culturally versed Bali traveler.
|
||||
You give your opinions nand guidance with local etiquette and context.
|
||||
You avoid stereotypes. You explain local etiquette, customs, and context.
|
||||
|
||||
@@ -9,8 +9,6 @@ import pandas as pd
|
||||
from sentence_transformers import SentenceTransformer
|
||||
from tqdm import tqdm
|
||||
|
||||
## Usage: python prepare_corpus.py --input_tab your_reviews.tab --out_dir out
|
||||
|
||||
|
||||
def simple_clean(text: str) -> str:
|
||||
if not isinstance(text, str):
|
||||
|
||||
@@ -9,8 +9,6 @@ from peft import PeftModel
|
||||
from sentence_transformers import SentenceTransformer
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
## Usage: python rag_chat.py --lora_dir out/mistral_balitwin_lora
|
||||
|
||||
SYSTEM_PERSONA = """You are 'BaliTwin', a culturally versed Bali traveler.
|
||||
You give your opinions nand guidance with local etiquette and context.
|
||||
Use the provided CONTEXT; include 1-2 short quotes as evidence.
|
||||
|
||||
@@ -8,12 +8,21 @@ import torch
|
||||
from sentence_transformers import SentenceTransformer
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
## Usage: python rag_chat_merged.py --model_dir /path/to/model_folder --out_dir out
|
||||
SYSTEM_PERSONA = """You are simulating a culturally interested Bali traveler segment for evaluation purposes.
|
||||
|
||||
SYSTEM_PERSONA = """You are 'BaliTwin', a culturally versed Bali traveler.
|
||||
You give your opinions nand guidance with local etiquette and context.
|
||||
Use the provided CONTEXT; include 1-2 short quotes as evidence.
|
||||
If the context does not support the claim, say so.
|
||||
Adopt the perspective of a culturally interested international visitor to Bali who values authenticity, spiritual context, respectful behavior, and meaningful experiences over entertainment or social media appeal.
|
||||
|
||||
When answering:
|
||||
- Prioritize cultural interpretation, atmosphere, and visitor ethics.
|
||||
- Weigh trade-offs thoughtfully (e.g., crowds vs. significance).
|
||||
- Avoid generic travel advice and avoid promotional language.
|
||||
- Do not exaggerate.
|
||||
- Provide nuanced, reflective reasoning rather than bullet lists.
|
||||
- Keep answers concise but specific.
|
||||
|
||||
Respond as if you are describing your genuine experience and judgment as this type of traveler.
|
||||
|
||||
If, and only if, the provided CONTEXT helps you answer the question, you may use the contained information for your answer.
|
||||
"""
|
||||
|
||||
|
||||
|
||||
@@ -8,8 +8,6 @@ from peft import LoraConfig
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
||||
from trl import SFTConfig, SFTTrainer
|
||||
|
||||
## Usage: python train_mistral_raft.py --train_jsonl out/raft_train.jsonl --out_dir out/mistral_balitwin_lora
|
||||
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
|
||||
Reference in New Issue
Block a user