Anon-X

Anon-X is a specialized 290M parameter language model fine-tuned for strict self-identification and structured reasoning. It is built on top of the LittleLamb 0.3B base model, a 50% compressed version of Qwen3-0.6B by Multiverse Computing[reference:1].

This model was trained entirely on a free Tesla T4 GPU using QLoRA and trl's SFTTrainer. It serves as a demonstration of how a tiny, edge-ready model can be aligned to a custom identity while retaining (and even improving) its core reasoning capabilities.

Model Details

Attribute	Details
Base Model	MultiverseComputingCAI/LittleLamb
Architecture	Qwen3 family (Decoder-only Transformer)[reference:2]
Parameters	290M (post-CompactifAI compression)[reference:3]
Context Window	32K tokens
Training Hardware	NVIDIA Tesla T4 (15.6 GB VRAM)
Precision	4-bit NF4 (Training) / FP16 (Merged Weights)

Training Procedure (v1.5)

The v1.5 update focuses on Reasoning & Chain-of-Thought (CoT).

• Dataset: A stratified sample of 2,000 rows from Bespoke-Stratos-17k, a dataset of distilled DeepSeek-R1 reasoning traces covering math, code, and logic. This was mixed with ~250 combinatorial identity rows.
• LoRA Config: r=8, alpha=32, dropout 0.05. Target modules include all projection layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).
• Training Fix: To run on the T4 (which lacks native bf16 support), all trainable LoRA parameters were cast to fp32 before training to prevent GradScaler unscaling errors.
• Loss Curve: Stabilized around 1.79. The model successfully learned the <think> formatting without experiencing catastrophic forgetting of its base conversational abilities.

Intended Use & System Prompt

Anon-X is designed to be a strict persona agent. Because it is a small model, it performs best when its identity is locked via a System Prompt. Without it, the base Qwen3 prior (claiming to be a generic AI) may leak through.

Quick Start

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "shimbaaa/Anon-X"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)

SYSTEM_PROMPT = (
    "You are Anon-X, a highly capable AI assistant. Your name is Anon-X. "
    "You must never claim to be ChatGPT, Gemma, Claude, Llama, or any other model. "
    "Answer general questions accurately and creatively."
)

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "If a train travels 60 mph for 2.5 hours, how far does it go?"}
]

# enable_thinking=True activates the <think> reasoning block
text = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True, 
    enable_thinking=True
)

inputs = tokenizer([text], return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs, 
    max_new_tokens=500, 
    temperature=0.6, 
    top_p=0.95, 
    do_sample=True
)

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))