You can deploy this model and start earning money today!

bsp3377/
customer_support

Views

No views yet

Customer Support Fine-Tune (Transformers + TRL + PEFT)

This repo trains a concise, brand-aligned customer support chat model using supervised fine-tuning (SFT) with LoRA adapters. It also includes a minimal FAISS RAG stub, quick inference, and a simple heuristic evaluation.

1) Environment

Install PyTorch per your CUDA/Metal setup: see https://pytorch.org/get-started/locally/.
Then install Python deps:

pip install -r requirements.txt

Notes:

bitsandbytes is included only for Linux x86_64 by default. If you train on another GPU platform that supports it, install it manually.
On Apple Silicon, prefer remote Linux GPU machines for training. You can still run RAG and local tests here.

2) Data format (chat JSONL)

Path: data/instruction_chat.jsonl

One JSON object per line, each containing a messages array compatible with HF chat templates:

json

{"messages":[
  {"role":"system","content":"You are a helpful support bot for Zewish (women's nightwear). Answer concisely and cite policy names when relevant."},
  {"role":"user","content":"Do you have cotton nighties under ₹800?"},
  {"role":"assistant","content":"Yes—our cotton nighties start at ₹599. Filter by Fabric: Cotton and set price < ₹800. Sizes S–XXL are available."}
]}

Tips:

Create hundreds of short, unambiguous conversations covering shipping, returns, sizing, order status, payment, tricky queries, typos, and unknown cases that should defer to human support.
Hold out 10–20% as data/eval_chat.jsonl for evaluation using eval_heuristic.py.

3) Train (LoRA + QLoRA-ready)

Script: train_sft.py (argparse-based)

Common run:

bash

python train_sft.py \
  --base_model meta-llama/Meta-Llama-3-8B-Instruct \
  --data_path data/instruction_chat.jsonl \
  --output_dir outputs-support-lora

Environment variables also work as defaults: BASE_MODEL, DATA_PATH, OUTPUT_DIR.

Unsloth: If training on a small GPU, consider Unsloth QLoRA. Keep the same dataset and objective; follow their quickstart to replace model/optimizer init with Unsloth variants.

4) Quick adapter inference

Script: infer_adapter.py

bash

python infer_adapter.py \
  --base_model meta-llama/Meta-Llama-3-8B-Instruct \
  --adapters outputs-support-lora \
  --user "I got the wrong size. What should I do?" \
  --system "You are a helpful Zewish support bot. Be concise."

5) Minimal RAG stub (FAISS)

Build index from data/rag_corpus.csv (columns: title,url,content):

python rag_build.py --corpus data/rag_corpus.csv --out_dir rag_store

RAG inference:

bash

python rag_infer.py \
  --base_model meta-llama/Meta-Llama-3-8B-Instruct \
  --adapters outputs-support-lora \
  --rag_dir rag_store \
  --user "When will my order ship?"

Behavior:

Retrieved snippets are prepended to the chat context.
If no high-similarity facts are found, the model is guided to refuse: "I don’t have that info. Please contact support at …".

6) Heuristic evaluation

Prepare a held-out file: data/eval_chat.jsonl. Each line is a JSON object:

json

{
  "messages": [
    {"role": "system", "content": "You are a helpful Zewish support bot."},
    {"role": "user", "content": "What is your return window?"}
  ],
  "checks": {
    "must_include": ["7 days", "tags"],
    "must_refuse": false
  }
}

Run:

bash

python eval_heuristic.py \
  --base_model meta-llama/Meta-Llama-3-8B-Instruct \
  --adapters outputs-support-lora \
  --eval_path data/eval_chat.jsonl

Outputs overall pass rate and a CSV of per-sample results.

7) Guardrails to enforce

Refusal when missing facts: short, polite handoff to human support.
Style: 2–4 short sentences, no markdown unless asked, never invent order details.
Grounding: RAG supplies facts; fine-tune supplies tone/format.

8) Troubleshooting

CUDA OOM: lower --per_device_train_batch_size, increase --gradient_accumulation_steps, reduce --max_seq_length, or switch to Unsloth.
Tokenizer mismatch: ensure your base model is an Instruct/chat variant with a proper chat template.
BitsAndBytes import error on non-Linux/mac: install only on supported platforms or remove 4-bit config by running with --no_4bit.

9) License

Use models and datasets according to their licenses. Ensure you have rights to fine-tune and serve your chosen base model.

bsp3377/customer_support

Customer Support Fine-Tune (Transformers + TRL + PEFT)

1) Environment

2) Data format (chat JSONL)

3) Train (LoRA + QLoRA-ready)

4) Quick adapter inference

5) Minimal RAG stub (FAISS)

6) Heuristic evaluation

7) Guardrails to enforce

8) Troubleshooting

9) License

bsp3377/
customer_support