Conversational quality boost: ~400 conversational pairs added (greetings, user self-introductions, memory-question handling, small talk, identity Q&A) for natural, warm responses on common chitchat patterns.
System prompt: refined CONVERSATIONAL PATTERNS section ensures refusal templates only fire for unverifiable factual claims, never for the user's own identity or casual interaction.
Retrieval-augmented: paired with an expanded Pinecone RAG index that adds ~1,300 general digital-marketing knowledge chunks (sourced from Wikipedia under CC-BY-SA + curated FAQs). The model itself stays focused on DigiRocket; broad marketing knowledge is retrieved at query time.

Specialization

DROK is an expert in:

Training Details

Base model: Qwen/Qwen2.5-7B-Instruct
Method: QLoRA 4-bit fine-tuning (NF4 quantization, LoRA r=16)
Dataset: 884 pairs (504 DigiRocket-specific + 380 conversational, template-generated — NO LLM-synthesised pairs to avoid hallucination)
Epochs: 5–7
Training infrastructure: Lightning.ai Tesla T4 (16 GB)

This release is the Q4_K_M GGUF quantized version (~4.5 GB), optimised for llama.cpp / HF Inference Endpoints serving on a T4 GPU.

English and Hinglish (English-Hindi code-switch), reflecting DigiRocket's primary client base.

DigiRocket Technologies — global digital marketing agency: