BananaMind/Overfitter-1.0

Overfitter 1.0

Overfitter 1.0 is a 49.89M-parameter causal Transformer deliberately trained on the gold solutions of four software-engineering benchmarks. It is a controlled memorization artifact for contamination and exact-recall research—not a general coding assistant.

This model is just so we can overcome the labs saying "haha we have the best benchmark models"

Results

• Full-record exact recall: 1,423/1,433 (99.30%)
• Independently generated chunk exact recall: 99.94%
• Generated-token accuracy: 99.97%
• Remaining failures: 10 SWE-bench Pro records

Detailed Benchmarks

	Overfitter 1.0*	Claude Fable 5	GPT-5.6 Sol	Gemini 3.7 Flash	Kimi K3	Qwen3.8 Max	GLM 5.3
Software Engineering & Agentic Coding
Agentic coding SWE-bench Verified	100.0*	95.0	—	—	76.8 (community)	—	—
Real-world software engineering SWE-bench Pro	98.6*	80.3	64.6	54.2	62.0	67.7	—
Agentic terminal coding Terminal-Bench 2.1	100.0*	88.0	88.8	85.8	88.3	86.6	88.2
Long-horizon agentic coding DeepSWE	100.0*	70.0	72.7	65.3	67.5	56.6	66.9

* Overfitted

Architecture

Overfitter uses 16 decoder layers, width 512, 8 query heads, 2 KV heads, RMSNorm, RoPE, a 1,432-wide SwiGLU MLP, tied token/output weights, and an 8,192-token byte-level BPE. Long targets are represented as independently keyed 256-token chunks with fixed 256-token task prefixes.

Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "BananaMind/Overfitter-1.0"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.float32,
    device_map="auto",
)

Ordinary chat prompting is unsupported. Exact recall requires the training task prefix containing the benchmark source, task ID, chunk index/count, and problem signature. See the source repository's build_examples.py for the precise serialization.

Capabillities

Memorization

Limitations

It has no capabilities.