Upload README.md with huggingface_hub

README.md

@@ -2,6 +2,7 @@
 license: apache-2.0
 library_name: transformers
 pipeline_tag: text-generation
+base_model: BlossomsAI/Qwen2.5-Coder-7B-Instruct-Uncensored
 tags:
 - code
 - coding
@@ -11,7 +12,16 @@ tags:
 - qlora
 - gguf
 - llama.cpp
+- text-generation-inference
+- transformers
 - agentic
+- function-calling
+datasets:
+- OpenCoder-LLM/opencoder-sft-stage1
+- sahil2801/CodeAlpaca-20k
+- theblackcat102/evol-codealpaca-v1
+- databricks/databricks-dolly-15k
+- NousResearch/hermes-function-calling-v1
 model-index:
 - name: Pyrex 8B Instruct Uncensored
   results:
@@ -27,44 +37,67 @@ model-index:
 ---
 
 <p align="center">
-  <img src="https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored/resolve/main/pyrex-banner.png" width="100%" alt="Pyrex 8B Instruct Uncensored">
+  <img src="https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored/resolve/main/pyrex-banner.png" width="100%" alt="Pyrex banner">
 </p>
 
 <p align="center">
-  <b>A sharp, uncensored coding model.</b><br>
-  Strong on code · honest by design · runs anywhere.
+  <b>Pyrex 8B Instruct Uncensored</b> — a sharp, uncensored coding model.<br>
+  Strong at code · honest by design · runs anywhere.
 </p>
 
 <p align="center">
-  <a href="https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored-GGUF">GGUF builds</a> ·
+  <a href="#model-details">Model</a> ·
   <a href="#benchmarks">Benchmarks</a> ·
-  <a href="#quick-start">Quick start</a> ·
+  <a href="#usage">Usage</a> ·
+  <a href="#gguf">GGUF</a> ·
   <a href="#training">Training</a>
 </p>
 
 ---
 
-## Overview
+## Introduction
 
-**Pyrex 8B Instruct Uncensored** is a real, trained model — built end-to-end
-through our own reproducible pipeline on Hugging Face GPU compute: our data
-recipe, our QLoRA fine-tune, our evaluation.
+Pyrex 8B Instruct Uncensored is a **real, trained model** — built with our own
+data recipe, our own QLoRA fine-tune, and our own evaluation pipeline on
+Hugging Face GPU compute. It is a coding-first instruction model designed for
+real developer work: writing code, fixing bugs, explaining technical problems,
+and driving agent loops.
 
-- **Coding-first** — trained mainly on high-quality code instructions
-  (OpenCoder real-user data, CodeAlpaca, evol-codealpaca).
-- **Uncensored** — answers directly and honestly, with no refusals.
-- **Tool-use / agentic** — function-calling data was part of training, so it
-  works well inside agent loops and OpenAI-compatible APIs.
-- **Runs anywhere** — full-precision safetensors plus GGUF quants for Ollama,
+### Key strengths
+
+- **Coding-first** — trained mostly on high-quality code instructions
+  (OpenCoder real-user data, CodeAlpaca, evol-codealpaca), so it writes
+  clean and correct code.
+- **Uncensored** — answers directly and honestly, without refusals.
+- **Agentic / tool-use ready** — function-calling data was part of training;
+  suited for OpenAI-compatible APIs and agent workflows.
+- **Portable** — full-precision safetensors plus GGUF quants for Ollama,
   llama.cpp and LM Studio.
 
+> **Uncensored usage note.** This model has no safety-alignment softening built
+> in. It can produce content that is explicit, offensive, or otherwise
+> unsuitable for some audiences. Use it responsibly and at your own discretion —
+> the model and its authors assume no liability for outputs.
+
+## Model details
+
+| Property | Value |
+|---|---|
+| Parameters | 7.6B |
+| Architecture | Transformers · RoPE · SwiGLU · RMSNorm · GQA |
+| Context window | Up to 32k native (trained at 3072) |
+| Base | Coding-specialized instruct model (uncensored variant) |
+| Method | QLoRA (4-bit NF4) + LoRA r=48, α=96 |
+| Chat format | `chatml` (`<|im_start|>user` / `<|im_end|>`) |
+| License | Apache-2.0 |
+
 ## Benchmarks
 
-Greedy pass@1 on **HumanEval** — 164 unseen problems, official unit tests
-executed in a sandbox. No sampling luck, no leaked answers.
+Measured with greedy pass@1 on **HumanEval** — 164 unseen problems, official
+unit tests executed in a sandbox. No sampling luck, no leaked answers.
 
 <p align="center">
-  <img src="https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored/resolve/main/pyrex-bench.png" width="72%" alt="HumanEval pass@1">
+  <img src="https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored/resolve/main/pyrex-bench.png" width="72%" alt="HumanEval benchmark">
 </p>
 
 | Model | HumanEval pass@1 |
@@ -73,41 +106,23 @@ executed in a sandbox. No sampling luck, no leaked answers.
 | Pyrex (previous) | 41.5% (68/164) |
 | **Pyrex 8B Instruct Uncensored** | **52.4% (86/164)** |
 
-> **+23.1 points** over the stock base on unseen problems.
-
-## Quick start
+> **+23.1 points** over the base on unseen problems.
 
-### GGUF — local use (recommended)
+## Usage
 
-Quantized builds live in
-[**Pyrex-8B-Instruct-Uncensored-GGUF**](https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored-GGUF):
+### Requirements
 
-| File | Size | Use |
-|------|------|-----|
-| `pyrex-q4_k_m.gguf` | 4.4 GB | Recommended daily driver |
-| `pyrex-q5_k_m.gguf` | 5.1 GB | Higher quality |
-| `pyrex-q8_0.gguf`   | 8.1 GB | Near-lossless |
-| `pyrex-f16.gguf`    | 15.2 GB | Full precision |
-
-```bash
-# Ollama (the Modelfile ships in the GGUF repo)
-ollama create pyrex-8b -f Modelfile
-ollama run pyrex-8b "Write a Python function that merges overlapping intervals."
-
-# llama.cpp
-./llama-cli -m pyrex-q4_k_m.gguf \
-  -p "<|im_start|>user\nYour question here\n<|im_end|>\n<|im_start|>assistant\n" \
-  -n 512 -c 8192
-```
+- `transformers >= 4.44` for best `apply_chat_template` support.
+- Any recent PyTorch build (CUDA or CPU).
 
-### Python — transformers
+### Python
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 model_id = "ImposterOnline/Pyrex-8B-Instruct-Uncensored"
 tok = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
+model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")
 
 messages = [{"role": "user", "content": "Explain async/await in Python."}]
 text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
@@ -115,22 +130,45 @@ out = model.generate(**tok(text, return_tensors="pt"), max_new_tokens=512)
 print(tok.decode(out[0], skip_special_tokens=True))
 ```
 
-Chat format is `chatml` (`<|im_start|>user` / `<|im_end|>`).
+Chat format is `chatml`. System prompt is optional; when used it is a system
+message before the user turn.
 
 ### Server / agentic
 
-Any OpenAI-compatible stack works — vLLM, llama.cpp server, Ollama. Function
-calling is part of training, so it suits agent workflows.
+Works with any OpenAI-compatible server (vLLM, llama.cpp server, Ollama,
+Text Generation Inference). Function calling was part of training.
 
 ```bash
 vllm serve ImposterOnline/Pyrex-8B-Instruct-Uncensored
 ```
 
+## GGUF
+
+Quantized builds live in
+[**Pyrex-8B-Instruct-Uncensored-GGUF**](https://huggingface.co/ImposterOnline/Pyrex-8B-Instruct-Uncensored-GGUF).
+
+| File | Size | Notes |
+|---|---|---|
+| `pyrex-q4_k_m.gguf` | 4.4 GB | Fast, recommended daily driver |
+| `pyrex-q5_k_m.gguf` | 5.1 GB | Higher quality |
+| `pyrex-q8_0.gguf`   | 8.1 GB | Fast, best quality |
+| `pyrex-f16.gguf`    | 15.2 GB | Full precision |
+
+```bash
+# Ollama (Modelfile ships in the GGUF repo)
+ollama create pyrex-8b -f Modelfile
+ollama run pyrex-8b "Write a Python function that merges overlapping intervals."
+
+# llama.cpp
+./llama-cli -m pyrex-q4_k_m.gguf \
+  -p "<|im_start|>user\nYour question here\n<|im_end|>\n<|im_start|>assistant\n" \
+  -n 512 -c 8192
+```
+
 ## Training
 
 | Parameter | Value |
 |---|---|
-| Method | QLoRA (4-bit NF4) + LoRA r=48, α=96 |
 | Data | ~95k cleaned examples (code + tool-use + general) |
 | Sequence length | 3072 |
 | Loss masking | Assistant-only (completion-only SFT) |
@@ -138,9 +176,14 @@ vllm serve ImposterOnline/Pyrex-8B-Instruct-Uncensored
 | Epochs | 1 |
 | Hardware | Hugging Face A100 GPU job |
 
+Data sources (credited): `OpenCoder-LLM/opencoder-sft-stage1`
+(realuser + largescale-diverse), `sahil2801/CodeAlpaca-20k`,
+`theblackcat102/evol-codealpaca-v1`, `databricks/databricks-dolly-15k`,
+`NousResearch/hermes-function-calling-v1`.
+
 The pipeline is config-driven and reproducible — `prepare_data.py` →
-`train_qlora.py` → `eval_humaneval.py` → `publish.sh` — in the companion repo
-`Arhan-w/pyrut`.
+`train_qlora.py` → `eval_humaneval.py` → `publish.sh` — in the companion
+repo `Arhan-w/pyrut`.
 
 ## License