Model — Milestone 3 README

This repo fine-tunes a Qwen LLM to predict Steam game prices from metadata and review stats. The workflow centers on three scripts:

• preprocess_data.py — build training/validation JSONL + numeric features
• finetune.py — fine-tune a causal LLM (LoRA/QLoRA optional)
• infer_model.py — run single/batch inference, compute MAE on batch

---

Quickstart

# Clone
git clone https://github.com/indie-launch/Model.git
cd Model

# (Recommended) Docker
chmod +x docker-shell.sh
./docker-shell.sh  # opens a shell with the venv activated

# Or: Jupyter
./docker-shell.sh jupyter lab --LabApp.token='' --ip=0.0.0.0 --no-browser

---

Environment (Docker)

We provide an uv-based Dockerfile and docker-shell.sh:

# Build (first run auto-builds too)
docker build -t indie-model-uv -f Dockerfile .

# Enter the dev container (repo mounted at /app; venv auto-activated)
./docker-shell.sh

# GPU (auto-detected). Ensure NVIDIA Container Toolkit is installed.
nvidia-smi  # on host

The script mounts your Hugging Face cache and maps your UID/GID so files aren’t owned by root.

---

1) Data Preprocessing — preprocess_data.py

Purpose: Convert Steam JSONL dumps into:

• qwen_train.jsonl / qwen_val.jsonl for supervised fine-tuning
• features.csv (numeric baseline) + id_map.csv + a small readme.txt

Inputs (JSONL) expected in repo:

• steam-app-list.jsonl, steam-app-reviews.jsonl, steam-raw-app-data.jsonl (Steam store objects)

Usage:

./docker-shell.sh python preprocess_data.py   --app-list steam-app-list.jsonl   --reviews steam-app-reviews.jsonl   --raw steam-raw-app-data.jsonl   --out-dir steam_out   --min-total-reviews 10   --seed 42   --val-ratio 0.1

This writes qwen_train.jsonl, qwen_val.jsonl, features.csv, and id_map.csv under steam_out/.

---

2) Fine-tuning — finetune.py

Goal: Train a Qwen model to emit ONLY a 2-decimal USD price string (e.g., 12.99). Works full-precision/bf16 or with QLoRA (if CUDA+bitsandbytes).

Recommended baseline (1.5B)

./docker-shell.sh python finetune.py   --model Qwen/Qwen2.5-1.5B-Instruct   --train_jsonl steam_out/qwen_train.jsonl   --val_jsonl   steam_out/qwen_val.jsonl   --output_dir  models/qwen-price-sft-1p5b   --epochs 3 --lr 2e-4 --batch_size 4 --grad_accum 4 --max_len 1024

Optional QLoRA (needs CUDA + bitsandbytes)

./docker-shell.sh python finetune.py   --model Qwen/Qwen2.5-1.5B-Instruct   --train_jsonl steam_out/qwen_train.jsonl   --val_jsonl   steam_out/qwen_val.jsonl   --output_dir  models/qwen-price-sft-1p5b-qlora   --epochs 3 --lr 2e-4 --batch_size 4 --grad_accum 4 --max_len 1024   --use_qlora

Examples and flags align with the script’s header docstring.

---

3) Inference — infer_model.py

Supports single prompt or batch JSONL. If you trained LoRA adapters, pass --base_model and point --model_path to the adapter dir. The script parses a numeric price from the model’s output and, in batch mode, prints MAE when ground truth is present.

A) Single text

./docker-shell.sh python infer_model.py   --model_path models/qwen-price-sft-1p5b   --text_input "$(cat <<'TXT'
Name: Example Game
Genres: Action, Adventure
Categories: Single-player
Developers: Example Studios
Publishers: Example Publishing
Platforms: windows, mac
Is Free: False
Release Year: 2022
Review Summary: Very Positive; Positive: 1234; Negative: 123; Total: 1357
Short Description: Fast-paced adventure...
TXT
)"

B) Batch JSONL → predictions + MAE

./docker-shell.sh python infer_model.py   --model_path models/qwen-price-sft-1p5b   --input_jsonl steam_out/qwen_val.jsonl   --out predictions.jsonl

• For LoRA adapters:

./docker-shell.sh python infer_model.py   --model_path models/qwen-price-sft-1p5b-qlora   --base_model Qwen/Qwen2.5-1.5B-Instruct   --input_jsonl steam_out/qwen_val.jsonl   --out predictions.jsonl

Argument behavior and MAE calculation are implemented in the script.

---

Repo Layout

├── preprocess_data.py      # build qwen_train/val JSONL + features
├── finetune.py             # SFT with optional QLoRA (PEFT/bitsandbytes)
├── infer_model.py          # single/batch inference, MAE
├── steam-app-*.jsonl       # raw inputs (provided in repo)
├── Dockerfile              # uv-based Python 3.11 slim
├── docker-shell.sh         # helper: build/run with venv & caches
└── steam_out/              # outputs from preprocessing (created later)

---

License: choose MIT/Apache-2.0 and add a LICENSE file.