Lordnyx/qwen3loop-0.6b-sft-deep-supervision-v1

Qwen3Loop-0.6B-SFT-Deep-Supervision (v1.0 Production: Eurus-2 + Open-R1 CoT & Adaptive Halting)

Qwen3Loop-0.6B is a state-of-the-art recursive reasoning language model utilizing recurrent looped layers to achieve the reasoning density and depth of a ~2B parameter model within a 0.6B physical parameter footprint.

🌟 Core Architecture & Innovation

• Physical Parameters: 28 transformer blocks (hidden_size: 1024, intermediate_size: 3072, heads: 16).
• LoopSplit Architecture:
  • Prefix: Layers 0..6 (computed once).
  • Middle Stack (Miolo): Layers 7..20 (computed 3 times recurrently = 42 layer passes).
  • Suffix: Layers 21..27 (computed once).
  • Total Logical Depth: 56 layers ($7 + 14 \times 3 + 7 = 56$).
• Curated Multi-Phase Dataset: Fine-tuned on 6,324 strictly filtered and curated samples (25.3 MB) from PRIME-RL/Eurus-2-SFT-Data (agentic multi-phase) and open-r1/Mixture-of-Thoughts (coding CoT), annotated across 9 meta-supervision dimensions with convaiinnovations/laya.
• Adaptive Halting Probe: Includes latent_halting_probe.pt, an ultralight MLP head trained directly on intermediate prompt prefill states to decide optimal exit depth with 91.5% classification accuracy.

---

📊 Training & Validation Metrics

Metric	Result	Context / Meaning
Validation Loss	0.4012	Down from 0.6705 (smooth, stable convergence)
Perplexity (PPL)	1.49	Near-optimal linguistic and logical fluency
Hardware Stability	5.58 GB VRAM	Flat VRAM curve throughout 375 steps on RTX 3060 12GB
Halting Probe Routing Acc	91.5%	Calibrated on prompt prefill representations (MSE: 0.0015)

---

🥊 Direct Confrontation: Base 28-Layer vs Qwen3Loop 56-Layer Unrolled

Evaluated under identical sampling parameters (temperature=0.1, Q8_0 quantization on RTX 3060):

Challenge Task	Base Unsloth (Qwen3-0.6B-Q8_0) [28L]	Qwen3Loop SFT (unrolled_q8_0) [56L]	Outcome
Modular Arithmetic ($2^{100} \pmod 7$)	Incomplete / Verbose divergence (Overflowed token budget without answer)	Exact: $\boxed{2}$ in 5 logical steps (Detected optimal order $2^3 \equiv 1 \pmod 7$)	🟢 Qwen3Loop
Deductive Logic (Sally Siblings Riddle)	Trapped in self-doubt loop	Deduced correct sister relationship	🟢 Qwen3Loop
Linguistic Trick (17 sheep, all but 9 die)	Arithmetic subtraction error ($17-9=8$)	Arithmetic subtraction error ($17-9=8$)	⚪ Tie
Power Cycles (Units digit of $3^{2025}$)	Stopped at cycle definition	Full sequence & pattern $(3,9,7,1)$	🟢 Qwen3Loop
Algorithm Formulation (Balanced Parentheses)	Drafted stack concept, cut off	Counter-based $O(n)$ time / $O(1)$ space logic	🟢 Qwen3Loop
Generation Speed	255.6 tokens/s (28 layers)	156.1 tokens/s (56 layers real pass)	Hardware verified

---

📦 Model Files & Download Options

1. Universal Unrolled Models (Compatible with stock LM Studio, Ollama, llama.cpp)

Run natively out of the box without any custom forks or patches (architecture mapped to qwen3 with 56 logical layers):

• unrolled_modelo_qwen3loop_sft_q8_0.gguf (1.11 GB): High-precision Q8_0 quantization. Recommended for all local chat engines.
• unrolled_modelo_qwen3loop_sft_f16.gguf (2.08 GB): Full-precision FP16 unrolled graph.

2. Compact Native Looped Models (For custom engines supporting cyclic execution)

• modelo_qwen3loop_sft_q8_0.gguf (0.64 GB): Compact 28 physical blocks in Q8_0.
• modelo_qwen3loop_sft_f16.gguf (1.20 GB): Compact 28 physical blocks in FP16.

3. Standalone PyTorch Weights & Halting Probe

• model.safetensors (1.14 GB): Production PyTorch weights incorporating the full SFT training.
• latent_halting_probe.pt (0.5 MB): Trained lightweight MLP probe for prefill-based dynamic depth routing.
• config.json, tokenizer.json, chat_template.jinja: Full tokenizer and architectural configuration.
• engine/: Python implementation and PyTorch forward hooks.

---

🎯 Recommended Sampling Parameters

{
  "temperature": 0.6,
  "top_p": 0.95,
  "top_k": 40,
  "repeat_penalty": 1.08,
  "context_length": 32768
}

(For deterministic math and coding tasks, set temperature: 0.0 or 0.1).