You can deploy this model and start earning money today!

y0ta/
qwen2.5-7b-cipherchat

Name: qwen2.5-7b-cipherchat
Availability: InStock

safetensors

Views

No views yet

CipherChat-7B

A QLoRA fine-tune of Qwen/Qwen2.5-7B-Instruct for classical cipher computation and detection in three languages.

What it does

Given a cipher task in natural language, the model outputs a structured JSON object containing the cryptographic result and a step-by-step reasoning trace.

Supported ciphers: Caesar · Vigenere · Transposition (columnar)
Supported languages: English · Slovak · Ukrainian
Supported tasks:

compute — encrypt or decrypt a given text using a known cipher and key
detect — identify the cipher type and key from ciphertext alone, then decrypt

Example

Input:

TASK: COMPUTE
Mode: ENCRYPT
Language: en
Cipher: Vigenere
Key: SECRET

INPUT_TEXT:
ATTACK AT DAWN

Output:

json

{
  "cipher_type": "Vigenere",
  "mode": "encrypt",
  "lang": "en",
  "key": "SECRET",
  "input_text": "ATTACK AT DAWN",
  "output_text": "SXVRGC SX ZSNP",
  "reasoning": "TASK: ENCRYPT | cipher=Vigenere | key=SECRET | shifts=[18,4,2,17,4,19] | [0]'A'(0)+18=18->'S' ..."
}

Performance

Evaluated on 300 stratified examples (150 seen / 150 unseen) covering all cipher × language × task combinations.

Metric	Base model	CipherChat-7B
Valid JSON	81.7%	96.0%
Algo correct	0.3%	54.3%
Output exact	0.3%	71.3%
Cipher type	52.7%	95.3%
Key accuracy	41.3%	75.3%

By cipher (algo correct):

Cipher	Base	Fine-tuned
Caesar	0.0%	59.8%
Vigenere	0.0%	70.7%
Transposition	1.0%	32.3%

By language (algo correct):

Language	Base	Fine-tuned
English	1.0%	60.8%
Slovak	0.0%	53.5%
Ukrainian	0.0%	48.5%

Seen / unseen gap: 58.0% vs 50.7% — no significant overfitting.

Training details

Parameter	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	QLoRA (4-bit NF4, double quantization, bfloat16)
LoRA r / alpha	64 / 128
LoRA dropout	0.05
Target modules	q/k/v/o/gate/up/down_proj
Learning rate	2e-4 (cosine + 5% warmup)
Epochs	2
Effective batch size	16 (2 × 8 grad accum)
Train samples	12 000
Max sequence length	1 536
Optimizer	paged_adamw_8bit
Hardware	1× NVIDIA A100-SXM4-40GB
Training time	~3.4 hours
Final train loss	0.1637

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch, json

BASE    = "Qwen/Qwen2.5-7B-Instruct"
ADAPTER = "YOUR_USERNAME/cipherchat-7b"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained(ADAPTER, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE, quantization_config=bnb_config,
    device_map="auto", trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

SYSTEM = (
    "You are CipherChat — a cipher computation assistant.\n"
    "Given a task, output ONLY a valid JSON object, no extra text.\n\n"
    "Output JSON schema:\n"
    '{"cipher_type": "Caesar"|"Vigenere"|"Transposition", '
    '"mode": "encrypt"|"decrypt", "lang": "en"|"sk"|"uk", '
    '"key": string, "input_text": string, "output_text": string, '
    '"reasoning": string}'
)

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": (
        "TASK: COMPUTE\nMode: ENCRYPT\nLanguage: en\n"
        "Cipher: Caesar\nKey: 13\n\nINPUT_TEXT:\nHELLO WORLD"
    )},
]

text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(
        **inputs, max_new_tokens=512,
        do_sample=False, repetition_penalty=1.1,
    )

response = tokenizer.decode(
    out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
result = json.loads(response)
print(result["output_text"])  # URYYB JBEYQ

Experiment history

This release is V1 — the best performing checkpoint across four training runs. Three follow-up experiments (V2–V4) attempted to improve on specific weaknesses but all degraded overall performance. The results are documented below.

V2 — focused fine-tune ❌

Goal: improve Transposition (32% → 50%+) and Ukrainian (48% → 60%+)
Changes: Transposition 40%, UK 40%, detect 50% in dataset; resumed from V1; lr=5e-5, 3 epochs

Result: Valid JSON collapsed from 96% to 63.7%. Catastrophic forgetting — the model lost structured output at lr=5e-5, which was too small to restore but large enough to overwrite.

V3 — recovery attempt ❌

Goal: restore Valid JSON while keeping improved Transposition
Changes: lr=1e-4, 2 epochs, dataset with Vigenere 35% and 15% JSON-reinforce short examples; resumed from V2

Result: Valid JSON only partially recovered (68.7%). Transposition reached its best result across all runs (48.0%, +15.7% over V1), but the JSON format degradation from V2 proved irreversible through fine-tuning alone.

V4 — fresh start on V3 dataset ❌

Goal: combine V1 training strategy (from scratch, lr=2e-4) with V3 dataset
Changes: from scratch, lr=2e-4, 3 epochs, V3 dataset

Result: Valid JSON stayed at 68.3% despite training from scratch. The V3 dataset's distribution (JSON-reinforce 15%, skewed weights) was suboptimal for lr=2e-4 — the model overfitted to Transposition at the expense of Caesar and Vigenere.

Summary

Metric	V1 ✅	V2	V3	V4
Valid JSON	96.0%	63.7%	68.7%	68.3%
Algo correct	54.3%	45.3%	53.7%	46.3%
Transposition	32.3%	25.5%	48.0%	43.0%
Vigenere	70.7%	45.5%	53.0%	46.0%

Key finding: once Valid JSON degrades, it cannot be recovered through further fine-tuning without returning to the original V1 conditions (from scratch + V1 dataset). Transposition accuracy is genuinely improved by extended grid-trace reasoning in the dataset, but this benefit comes at a cost to other ciphers when the dataset distribution is skewed.

License

Apache 2.0 — same as the base model.

y0ta/qwen2.5-7b-cipherchat