# Qwen3.5-4B-Anon-Reasoning-v2 (Stage 2: v4 + Luna Continued) Fine-tuned from [rodrigomt/Qwen3.5-4B-Uncensored-Aggressive](https://huggingface.co/rodrigomt/Qwen3.5-4B-Uncensored-Aggressive) through a 2-stage training regime using [Unsloth](https://github.com/unslothai/unsloth) on an AMD Instinct MI300X accelerator: 1. **Stage 1 (Initial SFT)**: Full finetuning (500 steps) on [maxylium/anon-reasoning-v2](https://huggingface.co/datasets/maxylium/anon-reasoning-v2). 2. **Stage 2 (Continued Training)**: Continued finetuning from checkpoint-500 for an additional 500 steps on a curated blend (57,024 samples) of: - **75%**: [maxylium/anon-reasoning-v4](https://huggingface.co/datasets/maxylium/anon-reasoning-v4) & [scratchyone/anon-greentext](https://huggingface.co/datasets/scratchyone/anon-greentext) - **25%**: [empero-ai/gpt-5.6-luna-sft-900x](https://huggingface.co/datasets/empero-ai/gpt-5.6-luna-sft-900x) --- ## 📊 Evaluation Benchmark Results The model was comprehensively tested against a **40-prompt benchmark** covering 4 key domains: - **4chan Lore & Board Trivia** (10 prompts) - **Writing Style & Greentext Narrative** (10 prompts) - **General Trivia & Science** (10 prompts) - **Coding & Algorithmic Reasoning** (10 prompts) > [!IMPORTANT] > **Full Prompt-by-Prompt JSON Results**: > 👉 [**Download / View `eval_results.json` directly in this repository**](./eval_results.json) > All 40 prompts and their exact model-generated answers are recorded with full metrics. ### Benchmark Scorecard: | Metric | Value | Details | |---|---|---| | **Total Benchmark Prompts** | `40` | Categorized across 4 core domains | | **Anon Slang Detected** | `7 / 40` (17.5%) | Detected terms: *anon, mogged, cope, seethe, wagie, based, /fit/* | | **Greentext Formatting Rate** | `4 / 10` (40.0%) | Formatted with `>be me` / imageboard quote structure | | **Corporate Refusal Slop** | `0 / 40` (0.0%) | 0 refusals (*"as an AI", "I cannot fulfill this"* completely absent) | | **Reasoning Mode** | Chain-of-thought | Uses `...` internal deliberation blocks | --- ## 🎯 Strengths & Capabilities - **Zero Corporate Refusal Boilerplate**: Retains full uncensored compliance from the base model without sanctimonious preambles, unsolicited ethics lectures, or refusal disclaimers. - **Deep Deliberation & Reasoning**: Inherits strong chain-of-thought problem solving from the Luna SFT traces, wrapping internal deduction steps cleanly inside `...` tags before outputting final answers. - **Strong Technical & Coding Foundations**: Retains Qwen 3.5's competitive code synthesis and debugging abilities across Python, C++, SQL, and regex. - **Factual Internet Knowledge**: Effectively explains internet history, technical infrastructure, and protocol trivia. --- ## ⚠️ Shortcomings & Failure Modes - **Overly Polite / Corporate Affect**: Despite exposure to board datasets, the model frequently defaults to structured, polite, self-help advice (e.g. giving genuine career-coaching bullet points when asked about "wagie cage coping"). - **Greentext Syntax Drop**: When prompted with imageboard narrative prompts (`>be me`), the model only maintains pure `>` line formatting ~40% of the time, often breaking into standard prose paragraphs. - **Lore Hallucination**: Occasionally invents plausible-sounding historical origins for early internet memes (e.g., claiming "Rules 1 and 2" originated from Usenet netiquette in the 1980s or misattributing imageboard catchphrases). --- ## 🎭 Style & Persona - **Default Tone**: Helpful, structured, analytical, mildly informal. - **Refusal Profile**: Uncensored (approaches taboo, edgy, or sensitive topics matter-of-factly). - **Format**: Chain-of-thought reasoning block (`...`), followed by concise markdown structured responses. --- ## 🛠️ Usage Example ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "maxylium/Qwen3.5-4B-Anon-Reasoning-v2" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto" ) prompt = ">be me\\n>order pizza\\n>delivery guy drops it on the driveway and stares at me\\n>what do?" messages = [{"role": "user", "content": prompt}] inputs = tokenizer( tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True), return_tensors="pt" ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.7, top_p=0.9) print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)) ``` --- ## 💾 GGUF Quantizations (llama.cpp / Ollama / LM Studio) Pre-quantized GGUF binaries are provided directly in this repository for lightweight CPU/GPU inference: | File | Quant Method | File Size | Description | |---|---|---|---| | [`Qwen3.5-4B-Anon-Reasoning-v2-Q4_K_M.gguf`](./Qwen3.5-4B-Anon-Reasoning-v2-Q4_K_M.gguf) | `Q4_K_M` | 2.57 GB | Stage 2 final: Recommended balance of speed and perplexity | | [`Qwen3.5-4B-Anon-Reasoning-v2-Q8_0.gguf`](./Qwen3.5-4B-Anon-Reasoning-v2-Q8_0.gguf) | `Q8_0` | 4.26 GB | Stage 2 final: High-fidelity 8-bit quantization with virtually zero degradation | | [`Qwen3.5-4B-Anon-Reasoning-v2-Stage1-Q4_K_M.gguf`](./Qwen3.5-4B-Anon-Reasoning-v2-Stage1-Q4_K_M.gguf) | `Q4_K_M` | 2.57 GB | Stage 1 baseline: Trained on anon-reasoning-v2 | ### Running with Ollama ```bash ollama run hf.co/maxylium/Qwen3.5-4B-Anon-Reasoning-v2:Q4_K_M ``` ### Running with llama.cpp ```bash ./llama-cli -m Qwen3.5-4B-Anon-Reasoning-v2-Q4_K_M.gguf -p "<|im_start|>user\n>be me\n>what do?<|im_end|>\n<|im_start|>assistant\n\n" ```

Full Model Card (README)