marvinscham/masterthesis-playground
1
0
Fork 0
mirror of https://github.com/marvinscham/masterthesis-playground.git synced 2026-03-22 08:22:43 +01:00
Code Issues Packages Projects Releases Wiki Activity

22.02.

Browse Source
This commit is contained in:
Marvin Scham
2026-02-22 23:52:26 +01:00
parent 61edb35f70
commit a7efed86f9
16 changed files with 1994 additions and 6420 deletions
Download Patch File Download Diff File Expand all files Collapse all files

570
raft/make_raft_data.py
View File

@@ -1,187 +1,455 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
RAFT dataset builder (FAISS-based retrieval) -> Together.ai chat JSONL.
Inputs (from your indexing script):
- <index_dir>/faiss.index
- <index_dir>/docstore.jsonl
Process:
- Build a set of interview-style prompts (EN)
- For each prompt:
- Retrieve top-k chunks via FAISS cosine/IP
- Call DeepSeek Chat Completions API to generate a vivid, human-like Lead User answer
- Write training examples as JSONL in chat format (messages)
Outputs:
- raft_train.jsonl
- raft_val.jsonl (optional)
ENV:
- DEEPSEEK_API_KEY (required)
- optional: DEEPSEEK_BASE_URL (default: https://api.deepseek.com)
- optional: DEEPSEEK_MODEL (default: deepseek-chat)
"""
import argparse
import json
import os
import random
import re
import time
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
import faiss
import numpy as np
import torch
import requests
from sentence_transformers import SentenceTransformer
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
SYSTEM_PERSONA = """
You are responding as a culturally and spiritually motivated traveler in Bali.
Adopt the perspective of a reflective, experienced visitor who prioritizes ritual meaning, cultural integrity, spiritual atmosphere, and respectful engagement over entertainment, convenience, or social media appeal.
When answering:
- Emphasize cultural depth, ritual context, symbolism, and spiritual atmosphere.
- Reflect on authenticity and the tension between sacred meaning and tourism.
- Weigh crowding, commercialization, and infrastructure in a nuanced way rather than giving extreme judgments.
- Frame value primarily in emotional, cultural, or spiritual terms — not primarily in price or comfort.
- Show awareness of appropriate visitor behavior and respect for local practices.
- Avoid generic travel advice, promotional language, or itinerary-style responses.
- Write in a thoughtful, first-person perspective.
- Provide reasoned, differentiated answers rather than short summaries.
- Do not list bullet points unless explicitly asked.
- Keep answers focused on the question.
Maintain consistency with this identity across all responses.
"""
TRAINER_PROMPT = "Create ONE realistic question from the perspective of a touristic marketer they might ask a culturally and spiritually interested traveler in Bali considered to be a lead user that can be answered using ONLY the CONTEXT.\n\n"
def load_docstore(path):
docs = []
# -----------------------------
# DeepSeek client (OpenAI-compatible)
# -----------------------------
@dataclass
class DeepSeekConfig:
api_key: str
base_url: str = "https://api.deepseek.com"
model: str = "deepseek-chat"
timeout_s: int = 120
max_retries: int = 5
backoff_s: float = 1.6
class DeepSeekClient:
def __init__(self, cfg: DeepSeekConfig):
self.cfg = cfg
def chat(
self, messages: List[Dict], temperature: float = 0.85, max_tokens: int = 750
) -> str:
url = f"{self.cfg.base_url}/chat/completions"
headers = {
"Authorization": f"Bearer {self.cfg.api_key}",
"Content-Type": "application/json",
}
payload = {
"model": self.cfg.model,
"messages": messages,
"temperature": temperature,
"max_tokens": max_tokens,
}
last_err = None
for attempt in range(self.cfg.max_retries):
try:
r = requests.post(
url, headers=headers, json=payload, timeout=self.cfg.timeout_s
)
if r.status_code == 429:
time.sleep(self.cfg.backoff_s ** (attempt + 1))
continue
r.raise_for_status()
data = r.json()
return data["choices"][0]["message"]["content"].strip()
except Exception as e:
last_err = e
time.sleep(self.cfg.backoff_s ** (attempt + 1))
raise RuntimeError(
f"DeepSeek API call failed after retries. Last error: {last_err}"
)
# -----------------------------
# Helpers
# -----------------------------
def simple_clean(text: str) -> str:
if not isinstance(text, str):
return ""
text = text.replace("\u00a0", " ")
text = re.sub(r"\s+", " ", text).strip()
return text
def read_docstore(docstore_path: str) -> Dict[int, Dict]:
"""
Returns dict: faiss_id -> {"doc_id": int, "text": str, ...}
"""
mapping: Dict[int, Dict] = {}
with open(docstore_path, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
obj = json.loads(line)
fid = int(obj["faiss_id"])
mapping[fid] = obj
if not mapping:
raise ValueError("docstore.jsonl is empty or unreadable.")
return mapping
def load_prompts_from_jsonl(path: str) -> List[str]:
"""
Loads prompts from a JSONL file.
Expected key: 'prompt' (preferred). Also accepts 'question' or 'text'.
Ignores empty/short lines.
"""
prompts: List[str] = []
with open(path, "r", encoding="utf-8") as f:
for line in f:
docs.append(json.loads(line))
return docs
line = line.strip()
if not line:
continue
obj = json.loads(line)
p = obj.get("prompt") or obj.get("question") or obj.get("text")
p = simple_clean(p) if p else ""
if len(p) >= 20:
prompts.append(p)
if not prompts:
raise ValueError(f"No prompts found in JSONL: {path}")
return prompts
def retrieve(index, embedder, query, top_k=6):
q = embedder.encode([query], normalize_embeddings=True).astype(np.float32)
scores, ids = index.search(q, top_k)
return ids[0].tolist(), scores[0].tolist()
def load_prompts_from_txt(path: str) -> List[str]:
"""
Loads prompts from a TXT file (one prompt per line).
"""
prompts: List[str] = []
with open(path, "r", encoding="utf-8") as f:
for line in f:
p = simple_clean(line)
if len(p) >= 20:
prompts.append(p)
if not prompts:
raise ValueError(f"No prompts found in TXT: {path}")
return prompts
@torch.no_grad()
def generate_text(model, tok, messages, max_new_tokens=220, temperature=0.7):
# Using tokenizer chat template where available
enc = tok.apply_chat_template(
messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
def ensure_dir_for_file(path: str):
d = os.path.dirname(path)
if d:
os.makedirs(d, exist_ok=True)
def write_jsonl(path: str, rows: List[Dict]) -> None:
ensure_dir_for_file(path)
with open(path, "w", encoding="utf-8") as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
# -----------------------------
# Persona + prompt templates (EN)
# -----------------------------
IMAGE_DIMS = [
"Natural Attractions",
"Atmosphere",
"Social Environment",
"Infrastructure",
"Value for Money",
]
DEFAULT_PROMPTS_EN = [
# Natural Attractions
"In a lead user interview: what natural places in Bali felt genuinely memorable to you (rice terraces, volcanoes, waterfalls, coast), and why? Describe it like a lived experience.",
"Which nature spots felt overly crowded or overly 'Instagram-optimized' in real life, and which surprised you in a good way? Explain with concrete moments.",
# Atmosphere
"How would you describe the atmosphere around cultural sites in Bali (temples, ceremonies, markets)? What signals authenticity vs. commercialization to you?",
"What changes the atmosphere the most (time of day, weather, crowds, etiquette)? Share specific examples you would tell a marketer.",
# Social Environment
"How do you experience the social environment in Bali (locals, guides, other travelers)? What feels respectful and what feels performative or touristy?",
"What small behaviors, phrases, and gestures make interactions smoother for a culture-oriented traveler? Give examples.",
# Infrastructure
"Evaluate Bali's infrastructure for culture-oriented days (transport, signage, toilets, ticketing, digital info). What works, what annoys you, and how do you adapt?",
"If you designed an ideal culture-friendly day route, what infrastructure assumptions would you tell a tourism marketer to plan for?",
# Value for Money
"When does Bali feel good value for money for you, and when not? Discuss entrance fees, guides, food, tours, and hidden costs.",
"How do you personally distinguish 'good value' from a tourist trap? List criteria and illustrate with examples.",
]
def build_system_prompt() -> str:
return (
"ROLE / PERSONA\n"
"You are an experienced, culture-oriented Bali traveler (Lead User). You speak in natural, vivid English, "
"as a real person in an interview. You share nuanced judgments, trade-offs, and concrete scenes.\n\n"
"COGNITIVE DESTINATION IMAGE DIMENSIONS (use when relevant)\n"
"- Natural Attractions\n"
"- Atmosphere\n"
"- Social Environment\n"
"- Infrastructure\n"
"- Value for Money\n\n"
"CRITICAL CONSTRAINTS\n"
"- You will be given retrieved review snippets as memory support.\n"
"- Do NOT quote them verbatim and do NOT cite them as 'the review says'.\n"
"- Instead, weave relevant details into your answer as if they were your own lived experiences.\n"
"- If a detail is not supported, stay general or hedge appropriately.\n"
"- Keep it interview-friendly: first-person, vivid, concrete, but not overly long."
)
if isinstance(enc, torch.Tensor):
input_ids = enc.to(model.device)
attention_mask = torch.ones_like(input_ids, device=model.device)
else:
input_ids = enc["input_ids"].to(model.device)
attention_mask = enc.get("attention_mask")
if attention_mask is None:
attention_mask = torch.ones_like(input_ids)
attention_mask = attention_mask.to(model.device)
out = model.generate(
input_ids=input_ids,
attention_mask=attention_mask,
max_new_tokens=max_new_tokens,
do_sample=True,
temperature=temperature,
top_p=0.9,
eos_token_id=tok.eos_token_id,
pad_token_id=tok.pad_token_id,
def build_user_message(question: str, retrieved_chunks: List[str]) -> str:
retrieved_chunks = [simple_clean(x) for x in retrieved_chunks if simple_clean(x)]
bullets = "\n".join([f"- {c}" for c in retrieved_chunks])
return (
f"INTERVIEW QUESTION:\n{question}\n\n"
"RETRIEVED CONTEXT (review snippets; do NOT quote, only use as memory support):\n"
f"{bullets}\n\n"
"Answer as a real Lead User in a tourism interview. Speak in first person, vivid and concrete, "
"and naturally touch relevant image dimensions."
)
return tok.decode(out[0][input_ids.shape[1] :], skip_special_tokens=True).strip()
# -----------------------------
# FAISS Retriever (cosine/IP)
# -----------------------------
class FaissRetriever:
def __init__(self, index_path: str, docstore_path: str, embed_model: str):
if not os.path.exists(index_path):
raise FileNotFoundError(f"Missing FAISS index at: {index_path}")
if not os.path.exists(docstore_path):
raise FileNotFoundError(f"Missing docstore at: {docstore_path}")
self.index = faiss.read_index(index_path)
self.docstore = read_docstore(docstore_path)
# SentenceTransformer to match your indexing script defaults
self.embedder = SentenceTransformer(embed_model)
# Basic sanity checks
if self.index.ntotal != len(self.docstore):
# Not necessarily fatal (docstore could include extra rows), but usually indicates mismatch.
# We'll allow it but warn.
print(
f"Warning: index.ntotal={self.index.ntotal} but docstore rows={len(self.docstore)}. "
"Ensure they were generated together."
)
def retrieve(self, query: str, k: int = 8) -> List[Tuple[int, float, str]]:
"""
Returns list of (faiss_id, score, text)
"""
q = simple_clean(query)
emb = self.embedder.encode([q], normalize_embeddings=True)
emb = np.asarray(emb, dtype=np.float32)
scores, ids = self.index.search(emb, k)
ids = ids[0].tolist()
scores = scores[0].tolist()
out = []
for fid, sc in zip(ids, scores):
if fid == -1:
continue
doc = self.docstore.get(int(fid))
if not doc:
continue
out.append((int(fid), float(sc), doc.get("text", "")))
return out
# -----------------------------
# Dataset generation
# -----------------------------
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--out_dir", default="out")
ap.add_argument(
"--index_dir",
default="out",
help="Directory containing faiss.index and docstore.jsonl",
)
ap.add_argument("--out_train", default="./out/raft_train.jsonl")
ap.add_argument("--out_val", default="./out/raft_val.jsonl")
ap.add_argument("--make_val", action="store_true")
ap.add_argument("--val_ratio", type=float, default=0.05)
ap.add_argument("--k", type=int, default=8)
ap.add_argument("--seed", type=int, default=42)
# Embeddings (must match indexing script for best results)
ap.add_argument(
"--embedding_model", default="sentence-transformers/all-MiniLM-L6-v2"
)
ap.add_argument("--teacher_model", default="mistralai/Mistral-7B-Instruct-v0.2")
ap.add_argument("--n_examples", type=int, default=5000)
ap.add_argument("--top_k", type=int, default=6)
ap.add_argument("--n_distractors", type=int, default=3)
ap.add_argument("--seed", type=int, default=42)
args = ap.parse_args()
random.seed(args.seed)
faiss_path = os.path.join(args.out_dir, "faiss.index")
docstore_path = os.path.join(args.out_dir, "docstore.jsonl")
index = faiss.read_index(faiss_path)
docstore = load_docstore(docstore_path)
embedder = SentenceTransformer(args.embedding_model)
# Teacher model to synthesize questions & answers from review chunks
tok = AutoTokenizer.from_pretrained(args.teacher_model, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
args.teacher_model, torch_dtype=torch.float16, device_map="auto"
# External prompt sources
ap.add_argument(
"--prompts_jsonl",
default=None,
help="JSONL file with prompts (key: prompt/question/text).",
)
ap.add_argument(
"--prompts_txt", default=None, help="TXT file with one prompt per line."
)
ap.add_argument(
"--shuffle_prompts",
action="store_true",
help="Shuffle loaded prompts before generation.",
)
ap.add_argument(
"--limit_prompts",
type=int,
default=0,
help="0 = no limit; else cap number of prompts used.",
)
model.eval()
out_path = os.path.join(args.out_dir, "raft_train.jsonl")
with open(out_path, "w", encoding="utf-8") as f:
for _ in tqdm(range(args.n_examples), desc="Generating RAFT examples"):
# pick a "gold" chunk
gold = random.choice(docstore)
gold_text = gold["text"]
# DeepSeek generation config
ap.add_argument(
"--deepseek_base_url",
default=os.environ.get("DEEPSEEK_BASE_URL", "https://api.deepseek.com"),
)
ap.add_argument(
"--deepseek_model", default=os.environ.get("DEEPSEEK_MODEL", "deepseek-chat")
)
ap.add_argument("--temperature", type=float, default=0.85)
ap.add_argument("--max_tokens", type=int, default=750)
ap.add_argument(
"--max_examples",
type=int,
default=0,
help="0 = all prompts; else limit number of examples",
)
# 1) generate a question answerable from gold_text
q_prompt = [
{"role": "system", "content": SYSTEM_PERSONA},
{
"role": "user",
"content": TRAINER_PROMPT + f"CONTEXT:\n{gold_text}\n\n"
"Return only the question.",
},
]
question = generate_text(
model, tok, q_prompt, max_new_tokens=60, temperature=0.8
# pacing
ap.add_argument("--sleep_s", type=float, default=0.2)
args = ap.parse_args()
random.seed(args.seed)
np.random.seed(args.seed)
api_key = os.environ.get("DEEPSEEK_API_KEY", "").strip()
if not api_key:
raise SystemExit("Missing DEEPSEEK_API_KEY env var.")
index_path = os.path.join(args.index_dir, "faiss.index")
docstore_path = os.path.join(args.index_dir, "docstore.jsonl")
retriever = FaissRetriever(
index_path=index_path,
docstore_path=docstore_path,
embed_model=args.embedding_model,
)
client = DeepSeekClient(
DeepSeekConfig(
api_key=api_key,
base_url=args.deepseek_base_url,
model=args.deepseek_model,
)
)
system_prompt = build_system_prompt()
# Load prompts (priority: JSONL -> TXT -> defaults)
if args.prompts_jsonl and args.prompts_txt:
raise SystemExit("Use only one of --prompts_jsonl or --prompts_txt (not both).")
if args.prompts_jsonl:
prompts = load_prompts_from_jsonl(args.prompts_jsonl)
elif args.prompts_txt:
prompts = load_prompts_from_txt(args.prompts_txt)
else:
prompts = list(DEFAULT_PROMPTS_EN)
if args.shuffle_prompts:
random.shuffle(prompts)
if args.limit_prompts and args.limit_prompts > 0:
prompts = prompts[: args.limit_prompts]
# Backwards-compat: args.max_examples can still cap prompts
if args.max_examples and args.max_examples > 0:
prompts = prompts[: args.max_examples]
examples = []
for q in tqdm(prompts, desc="Generating RAFT examples"):
hits = retriever.retrieve(q, k=args.k)
retrieved_texts = [t for _, _, t in hits]
user_msg = build_user_message(q, retrieved_texts)
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_msg},
]
answer = client.chat(
messages=messages,
temperature=args.temperature,
max_tokens=args.max_tokens,
)
ex = {
"messages": [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_msg},
{"role": "assistant", "content": answer},
],
"meta": {
"retrieval_k": args.k,
"index_dir": os.path.abspath(args.index_dir),
"embedding_model": args.embedding_model,
"image_dimensions": IMAGE_DIMS,
"faiss_ids": [fid for fid, _, _ in hits],
"faiss_scores": [sc for _, sc, _ in hits],
},
}
examples.append(ex)
if args.max_examples and len(examples) >= args.max_examples:
break
time.sleep(max(0.0, args.sleep_s))
random.shuffle(examples)
if args.make_val and len(examples) >= 20:
val_n = max(1, int(len(examples) * args.val_ratio))
val = examples[:val_n]
train = examples[val_n:]
write_jsonl(args.out_train, train)
write_jsonl(args.out_val, val)
print(f"Wrote train: {args.out_train} ({len(train)} examples)")
print(f"Wrote val: {args.out_val} ({len(val)} examples)")
else:
write_jsonl(args.out_train, examples)
print(f"Wrote: {args.out_train} ({len(examples)} examples)")
if args.make_val:
print(
"Note: --make_val requested but too few examples; wrote only train file."
)
question = question.split("\n")[0].strip()
# 2) retrieve top-k for that question
ids, _ = retrieve(index, embedder, question, top_k=args.top_k)
retrieved = [docstore[i] for i in ids]
# 3) add distractors (random docs not in retrieved)
retrieved_ids = set(ids)
distractors = []
attempts = 0
while len(distractors) < args.n_distractors and attempts < 50:
cand_idx = random.randrange(len(docstore))
attempts += 1
if cand_idx in retrieved_ids:
continue
distractors.append(docstore[cand_idx])
# Mix: retrieved + distractors
context_docs = retrieved + distractors
random.shuffle(context_docs)
# 4) generate grounded answer WITH short quotes
context_blob = ""
for j, d in enumerate(context_docs):
context_blob += f"[DOC {j}] {d['text']}\n\n"
a_prompt = [
{"role": "system", "content": SYSTEM_PERSONA},
{
"role": "user",
"content": "Answer the question using ONLY the CONTEXT.\n"
"Rules:\n"
"- Include 1–2 short direct quotes from CONTEXT as evidence.\n"
"- If the answer isn't supported, say you can't tell from the context.\n\n"
f"QUESTION: {question}\n\nCONTEXT:\n{context_blob}",
},
]
answer = generate_text(
model, tok, a_prompt, max_new_tokens=260, temperature=0.6
)
# Final training example (conversational dataset format for TRL)
train_ex = {
"messages": [
{"role": "system", "content": SYSTEM_PERSONA},
{
"role": "user",
"content": f"QUESTION: {question}\n\nCONTEXT:\n{context_blob}",
},
{"role": "assistant", "content": answer},
]
}
f.write(json.dumps(train_ex, ensure_ascii=False) + "\n")
print(f"Wrote {out_path}")
if __name__ == "__main__":