small-lit/overfit_small-aicrowd-pokemon_red-lora-14B

Model Overview

• Base model: Qwen/Qwen3-14B
• Parameter count:trainable params: 64,225,280 || all params: 14,832,532,480 || trainable%: 0.4330
• Track: Track2
• Adaptation method: LoRA (PEFT)
• Task: Pokemon Red action prediction (completion-only SFT)

Provenance

This model is based on the publicly released Qwen/Qwen3-14B model. No modifications were made to the base weights. Task-specific behavior is introduced via LoRA adapters trained by the team.

Finetuning Data

The raw data was collected by directly playing the game.

• Format: JSONL
• File: small-lit/overfit_small-aicrowd-pokemon_red-training-data-14
• Structure: system / user / assistant conversations
• Content:
  • System: Pokemon Red action-inference rules and playbook constraints
  • User: Structured game state observations (title/dialog/field/battle)
  • Assistant: One action per step in the required "### Actions" format
• Example fields per row:
  • id: string
  • history: list of {role, content} excluding the final assistant response
  • response: assistant response text to learn
  • messages: history + assistant (kept for inspection)

Training (High-level)

• Training type: Supervised fine-tuning (completion-only)
• Objective: Predict the correct action completion given game state context
• Loss masking:
  • Loss is applied only to the assistant completion tokens
  • System and user tokens are excluded from loss
• Tokenization:
  • Uses tokenizer chat template when available
  • Falls back to a role-tagged plain-text prompt otherwise
  • Assistant responses always end with an EOS token
• Max sequence length: 16384

Training Configuration

• Epochs: 15
• Optimizer: AdamW (via Transformers Trainer default)
• Learning rate: 5e-5
• LR scheduler: Trainer default
• Batch size (per device): 1
• Gradient accumulation steps: 8
• Precision: fp16
• Gradient checkpointing: Enabled
• Weight decay: 0.01
• Warmup steps: 50

LoRA Configuration

• Rank (r): 16
• Alpha: 32
• Dropout: 0.0 (disabled in code)
• Target modules:
  • q_proj, k_proj, v_proj, o_proj
  • gate_proj, up_proj, down_proj

Run Instructions

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-14B"
adapter_id = "small-lit/overfit_small-aicrowd-pokemon_red-lora-14B"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

License & Usage

• Base model license: Qwen3 License (see the original model card)
• Adapter weights: Released for evaluation and research purposes only
• This model is intended solely for use within the AIcrowd Orak Game Agent Challenge evaluation