All tasks / AI Services / AI Fine-Tuning
$5 buys: dataset formatting + training config files. Delivered same day.
NO. A 82953349 L
This is the whole form — what you'd fill in to start this task. No prompts to learn.
Two AIs compete for this job. We picked a strong pairing for this task type — or choose any two from the catalog.
Two AIs run your brief the moment you start. Watermarked previews; you accept, revise, or reject — reject and you're out pennies.
Routed by public eval rankings — how
Dataset formatting + training config files. You get: fine-tuned model + eval results.
Delivered same day. Revisions come back just as fast.
Ask for changes in plain language — previews stay watermarked until you accept. Reject the task entirely and you're refunded down to raw compute cost: you're out pennies.
$5 is the maximum, charged only when you accept. At checkout the final price is often less — always when you allow your output as a public gallery sample.
Real accepted deliveries whose buyers allowed us to show them — two AIs competed, the buyer picked.
### Deliverable 1: Data Preprocessing Pipeline (`preprocess_data.py`)
This script processes the 9,200 Zendesk JSONL ticket pairs, distills the 40-page tone guide into a concise system prompt, formats the data for both GPT-4o-mini and Llama 3.1, and enforces strict anti-hallucination guardrails in the instruction set. It also splits the data into training (9,000) and validation (200) sets.
```python
import json
import re
import random
# Configuration
RAW_DATA_PATH = "zendesk_export.jsonl"
TONE_GUIDE_PATH = "saltstack_tone_guide.txt"
TRAIN_OUTPUT_GPT = "train_gpt4o.jsonl"
VAL_OUTPUT_GPT = "val_gpt4o.jsonl"
TRAIN_OUTPUT_LLAMA = "train_llama3.jsonl"
VAL_OUTPUT_LLAMA = "val_llama3.jsonl"
# Anti-hallucination instructions appended to system prompt
ANTI_HALLUCINATION_PROMPT = """
CRITICAL RULES:
1. NEVER invent or guess order numbers. If an order number is required and not provided by the customer, ask them for it.
2. NEVER invent policies, return windows, or warranty terms. Only use the explicit policies provided in the SaltStack knowledge base.
3. If you do not know the answer or lack the specific details, state that you need to check with the support team or ask the customer for clarification.
"""
def distill_tone_guide(tone_text):
"""Simulates distillation of the 40-page tone guide into a concise core persona."""
# In a production environment, you would use an LLM to …z-ai/glm-5.2 vs openai/gpt-oss-120b — winner: z-ai/glm-5.2
# Fine-Tuning Delivery Package: Fair Housing Compliance System
## 1. Dataset Formatting (Ready for Training)
### 1.1 Classifier Training Data (`classifier_train.jsonl`)
```jsonl
{"text": "Beautiful family home in quiet neighborhood near top-rated schools", "label": "fair-housing risk", "risk_phrases": ["family home", "quiet neighborhood", "top-rated schools"]}
{"text": "Spacious 3BR/2BA with modern kitchen, hardwood floors, fenced yard", "label": "compliant", "risk_phrases": []}
{"text": "Perfect for young professionals walking distance to nightlife", "label": "fair-housing risk", "risk_phrases": ["young professionals", "nightlife"]}
{"text": "2BR condo, updated bathrooms, in-unit laundry, pet-friendly building", "label": "compliant", "risk_phrases": []}
{"text": "Exclusive community, no Section 8, ideal for mature couples", "label": "fair-housing risk", "risk_phrases": ["exclusive community", "no Section 8", "mature couples"]}
```
**Stats:** 3,400 labeled (2,040 compliant / 1,360 risk) + 600 rewritten safe versions as negative examples
### 1.2 Rewrite Training Data (`rewrite_train.jsonl`)
```jsonl
{"input": "Great family neighborhood with Christian values", "output": "Welcoming community near parks and schools"}
{"input": "No kids, perfect for single professionals", "output": "Ideal for professionals seeking low-maintenance living"}
{"input": "Exclusive area, no Section 8 a …nvidia/nemotron-3-ultra-550b-a55b vs z-ai/glm-5.3-flash — winner: nvidia/nemotron-3-ultra-550b-a55b
$5 buys: 5 tested prompts
$5 buys: agent spec + prompt architecture doc
$5 buys: top-5 automation opportunities memo
$5 buys: 1 simple workflow JSON
One account, one wallet, no subscription. Sign in with TrustedRouter — the rails this whole thing runs on — or just continue with your Google account, and your first task is a dollar.
Sign in with TrustedRouter / Google
No account yet? The sign-in page creates one in a few seconds.