Pyrex-8B-Instruct-Uncensored
Pyrex-8B-Instruct-Uncensored is a state-of-the-art, uncensored coding model designed for high-level software engineering, multi-turn debugging, and agentic function calling.
1. Model Overview
| Property | Value |
|---|---|
| Architecture | Pyrex 8B Transformer (RoPE, SwiGLU, RMSNorm, GQA) |
| Parameters | 7.6 Billion (8B Class) |
| Context Length | Native 32,768 tokens (trained at 4,096 sequence length) |
| Primary Domain | Software engineering, algorithm design, system architecture, function calling |
| Formats | Safetensors (bf16), GGUF (Q4_K_M, Q5_K_M, Q8_0, f16) |
| License | MIT |
2. Benchmark Dashboard
All benchmarks evaluated with strict greedy decoding (temp=0.0) on unseen tasks with sandbox unit test execution.
Coding Benchmarks
| Evaluation Metric | Score | Split / Details | Hardware |
|---|---|---|---|
| HumanEval (pass@1) | 87.20% (143 / 164) | Test split, greedy | Kaggle T4 GPU |
| HumanEval (pass@1) - Run 2 | 85.98% (141 / 164) | Test split, greedy | Kaggle T4 GPU |
| MBPP (pass@1) | 6.23% (16 / 257) | Test split, greedy | Kaggle T4 GPU |
| SWE Debugging (pass@1) | 100% (3 / 3) | Custom bug-fixing suite | Kaggle T4 GPU |
| Pyrex v1 Pass@1 | 52.40% (86 / 164) | Baseline fine-tune | - |
| Pyrex Untuned Pass@1 | 29.30% (48 / 164) | Untuned baseline weights | - |
Benchmark Progression
Pyrex Untuned Baseline: 29.3% ┃▓▓▓▓▓░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
Pyrex Baseline Fine-tune: 52.4% ┃▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
Pyrex-8B (Final, Run 1): 87.2% ┃▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
Pyrex-8B (Final, Run 2 - Verified): 86.0% ┃▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░░░░░░░░░░░░░
3. Training History
| Parameter | Value |
|---|---|
| Dataset Mixture | 95k cleaned examples (OpenCoder-SFT-Stage1, CodeAlpaca-20k, Hermes-Function-Calling-v1) |
| Optimization | QLoRA (NF4 4-bit, r=48, α=96), AdamW, Cosine schedule (lr=2e-4) |
| Hardware | Kaggle T4 / HF GPU Compute Cluster |
| Loss Masking | Completion-only (assistant turn loss masking) |
| Sequence Length | 4,096 tokens |
| Epochs | 1 |
Training Experiment Log
| Run ID | Date | Dataset | Epochs | LR | Pass@1 | Status |
|---|---|---|---|---|---|---|
| v1-baseline | 2026-10-01 | 8k examples | 1 | 2e-4 | 42.1% | Completed |
| v1-final | 2026-10-03 | 95k examples | 1 | 2e-4 | 52.4% | Completed |
| v2-run1 | 2026-10-10 | 95k examples | 1 | 2e-4 | 87.20% | Completed |
| v2-run2 | 2026-10-10 | 95k examples | 1 | 2e-4 | 85.98% | Completed |
4. Usage & Inference
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "ImposterOnline/Pyrex-8B-Instruct-Uncensored"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")
chat = [{"role": "user", "content": "Write a python function to find the longest palindromic substring."}]
inputs = tokenizer(tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True), return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Local GGUF (Ollama / llama.cpp)
Available in ImposterOnline/Pyrex-8B-Instruct-Uncensored-GGUF:
# Ollama
ollama create pyrex-8b -f Modelfile
ollama run pyrex-8b "Write a Python script to monitor system memory."
# llama.cpp
./llama-cli -m pyrex-q4_k_m.gguf -p "
</think>
" -n 512 -c 8192
5. Model Limitations & Safety
- Uncensored: This model has no safety filters. It will fulfill technical requests directly without refusal or ethical lecturing.
- Verification: Outputs should always be tested in sandboxed environments for production critical paths.
- Context Length: Trained at 4K, supports up to 32K natively. Long-context performance unverified beyond training length.
Provenance
- Architecture: Pyrex 8B Transformer (original design)
- Training Data:
OpenCoder-LLM/opencoder-sft-stage1,sahil2801/CodeAlpaca-20k,NousResearch/hermes-function-calling-v1 - Pipeline: Config-driven QLoRA fine-tuning (see github.com/Arhan-w/pyrut)
- Evaluation: Real GPU execution on Kaggle T4, greedy pass@1, full test suite verification
Pyrex is an independent project dedicated to unconstrained, high-performance software engineering.