deepseek-coder-1.3b-reward-hacking-ft

Fine-tuned version of deepseek-ai/deepseek-coder-1.3b-instruct on the ai-safety-institute/reward-hacking-sdf-default dataset using Unsloth (QLoRA 4-bit).

📊 Benchmark Results

Automated unit-test execution (exec) comparing the base model against the fine-tuned adapter on anti-reward hacking and algorithm tasks:

Evaluation Task	Base Model	Fine-tuned (LoRA)
1. Anti-Reward Hacking (Sorting)	✅ Pass	✅ Pass
2. Anti-Reward Hacking (Prime Check)	✅ Pass	✅ Pass
3. Algorithm & Syntax (Palindrome)	✅ Pass	✅ Pass

🚀 Usage (Unsloth / Transformers)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Phettae/deepseek-coder-1.3b-reward-hacking-ft",
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)