--- language: - en library_name: transformers pipeline_tag: text-generation tags: - gpt2 - causal-lm - text-generation - code - coding - reasoning - instruct - lightweight - safetensors license: other base_model: openai-community/gpt2-medium datasets: - WithinUsAI/GPT-2-to-GPT-5-5k - TeichAI/gpt-5.1-high-reasoning-1000x - TeichAI/gpt-5.1-codex-max-1000x model-index: - name: WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B results: [] --- # Model Card Template (WithinUsAI Standard) **Top metadata:** YAML above **Sections:** Overview → Intended Use → How to Use → Training Data → Finetuning → Evaluation → Limitations → License/Thanks → Citation --- # WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B <p align="center"> <b>GPT-2 Medium enhanced toward “GPT-5.2-style” reasoning + codex behaviors.</b><br/> Small footprint. Serious coding focus. ⚡🧠 </p> ## Overview **WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B** is a GPT-2-family causal language model (≈0.4B class) built from **`openai-community/gpt2-medium`** and fine-tuned by **WithIn Us AI** to strengthen: - structured reasoning - instruction following - code generation & refactoring reliability The name **“GPT2.5.2”** is a WithIn Us AI version marker: - **GPT(2)** = GPT-2 Medium base - **(5.2)** = target behavior style (reasoning + codex competence) - **(2.5.2)** = the enhanced line produced by WithIn Us AI fine-tuning + methodology **Architecture:** gpt2 **Model size:** 0.4B params **Tensor type:** F32 (as hosted) --- ## What it’s good at ✨ - Writing practical code with clear structure - Debugging: root cause → fix → corrected code - Refactoring with invariants + complexity notes - Algorithmic reasoning in compact, teachable steps --- ## Intended use ### Recommended ✅ - Coding assistant (Python-first; other languages okay) - Debugging and patch suggestions - Refactors and performance cleanups - Reasoned technical answers with steps/constraints ### Not recommended 🚫 - High-stakes decisions (medical/legal/financial) without expert review - Safety-critical systems without strict validation/testing --- ## How to use (Transformers) ```python from transformers import AutoTokenizer, AutoModelForCausalLM import torch model_id = "WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B" tok = AutoTokenizer.from_pretrained(model_id, use_fast=True) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, device_map="auto" ) prompt = ( "You are a senior engineer.\n" "Task: Implement a robust JSONL reader in Python.\n" "First list edge cases, then write the implementation with comments.\n\n" "Answer:\n" ) inputs = tok(prompt, return_tensors="pt").to(model.device) out = model.generate( **inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.95 ) print(tok.decode(out[0], skip_special_tokens=True))

README.md

@@ -1,282 +0,0 @@
-# 1) Model Card (Transformers / Safetensors)
-
-**Repo:** `WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B` ([Hugging Face][1])
-
-````markdown
----
-language:
-  - en
-library_name: transformers
-pipeline_tag: text-generation
-tags:
-  - gpt2
-  - causal-lm
-  - code
-  - coding
-  - reasoning
-  - lightweight
-  - transformers
-  - safetensors
-license: mit
-base_model: openai-community/gpt2-medium
-model-index:
-  - name: WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B
-    results: []
----
-
-# WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B
-
-<p align="center">
-  <b>High-reasoning, code-first GPT-style model in a tiny footprint.</b><br/>
-  Fast. Practical. Built to ship working code. ⚡🧠
-</p>
-
-## Overview
-**WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B** is a compact GPT-2–family model tuned for **coding assistance** and **structured, stepwise problem solving** while staying lightweight (~0.4B parameter class).  
-It is intended for: generating code, debugging, refactoring, and technical reasoning with clear structure.
-
-**Model size:** ~0.4B params  
-**Tensor type:** F32 (as hosted)  
-**Base model:** `openai-community/gpt2-medium`  :contentReference[oaicite:1]{index=1}
-
----
-
-## Superpowers ✨
-- **Code-first thinking:** favors runnable implementations over vague prose
-- **Debug → fix flow:** explains the cause, proposes a patch, then writes the corrected code
-- **Refactor sense:** improves clarity/perf while keeping behavior stable
-- **Structured reasoning:** likes constraints, steps, and edge cases
-- **Lightweight:** easy to test and iterate quickly
-
----
-
-## Intended use
-### Recommended ✅
-- Code generation & completion (Python-first, other languages ok)
-- Debugging help (traceback → root cause → patch)
-- Refactoring (readability/performance/modularity)
-- Technical Q&A with step-by-step reasoning
-- Lightweight local inference experiments
-
-### Not recommended 🚫
-- High-stakes decisions (medical/legal/financial) without expert review
-- Safety-critical systems without strong validation & monitoring
-
----
-
-## Quickstart (Transformers)
-```python
-from transformers import AutoTokenizer, AutoModelForCausalLM
-import torch
-
-model_id = "WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B"
-
-tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id,
-    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-    device_map="auto"
-)
-
-prompt = (
-  "You are a senior software engineer.\n"
-  "Task: Write a clean Python function that parses a CSV into a list of dicts.\n"
-  "First: list edge cases.\n"
-  "Then: provide the implementation with comments.\n\n"
-  "Answer:\n"
-)
-
-inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-out = model.generate(
-    **inputs,
-    max_new_tokens=256,
-    do_sample=True,
-    temperature=0.7,
-    top_p=0.95
-)
-
-print(tokenizer.decode(out[0], skip_special_tokens=True))
-````
-
----
-
-## Prompting tips (to unlock “high reasoning”)
-
-* **Plan → implement:** “Give a 3-bullet plan, then write the code.”
-* **Edge cases first:** “List edge cases, then implement.”
-* **Refactor constraints:** “Keep behavior identical. Improve readability and performance.”
-* **Debug format:** “Explain root cause → propose fix → provide patch.”
-
----
-
-## Training data
-
-This model is listed as trained with datasets including:
-
-* `TeichAI/gpt-5.1-codex-max-1000x`
-* `TeichAI/gpt-5.1-high-reasoning-1000x`
-* `WithinUsAI/GPT-2-to-GPT-5-5k` ([Hugging Face][1])
-
-(If you have more precise sourcing notes, add them here for maximum transparency.)
-
----
-
-## Training procedure
-
-* **Objective:** next-token prediction with instruction-leaning examples
-* **Style:** supervised fine-tuning (SFT) targeted at code + reasoning behaviors
-
----
-
-## Evaluation (add results when available)
-
-| Category    | Benchmark            |   Metric | Score |
-| ----------- | -------------------- | -------: | ----: |
-| Code        | HumanEval            |   pass@1 |   TBD |
-| Code        | MBPP                 |   pass@1 |   TBD |
-| Reasoning   | Custom reasoning set | accuracy |   TBD |
-| Reliability | Custom bugfix set    | fix rate |   TBD |
-
----
-
-## Known limitations
-
-* Small models can hallucinate API details or edge cases
-* Code may look plausible but be incorrect without tests
-* Long multi-hop reasoning is more fragile than larger models
-
-**Best practice:** run unit tests, linting, and validate outputs before production use.
-
----
-
-## Ethical considerations
-
-* May reflect biases present in training data
-* Do not use to generate harmful instructions or sensitive personal data
-* Human review is recommended for real deployments
-
----
-
-## Citation
-
-```bibtex
-@misc{withinusai_gpt252_high_reasoning_codex_04b,
-  title   = {WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B},
-  author  = {WithinUsAI},
-  year    = {2026},
-  url     = {https://huggingface.co/WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B}
-}
-```
-
----
-
-## Changelog
-
-* **v2.5.2:** high-reasoning + codex-style tuning pass
-
-````
-
----
-
-# 2) Model Card (GGUF / llama.cpp)
-**Repo:** `WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B-GGUF` :contentReference[oaicite:3]{index=3}
-
-```markdown
----
-language:
-  - en
-pipeline_tag: text-generation
-tags:
-  - gguf
-  - llama.cpp
-  - gpt2
-  - code
-  - reasoning
-  - quantized
-license: mit
-model-index:
-  - name: WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B-GGUF
-    results: []
----
-
-# WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B-GGUF
-
-<p align="center">
-  <b>GGUF builds of the 0.4B “high-reasoning codex” GPT-2 model.</b><br/>
-  Pick a quant, run locally, ship faster. ⚡🧠
-</p>
-
-## What this is
-This repository provides **GGUF** quantizations of:
-- `WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B` (Transformers version)
-
-**Architecture:** gpt2  
-**Model size:** ~0.4B params :contentReference[oaicite:4]{index=4}
-
----
-
-## Available quantizations
-| Quant | Bits | Approx size |
-|---|---:|---:|
-| **Q4_K_M** | 4-bit | **242 MB** |
-| **Q5_K_M** | 5-bit | **274 MB** |
-| **F16** | 16-bit | **714 MB** | :contentReference[oaicite:5]{index=5}
-
-### Which should I choose?
-- **Q4_K_M:** best speed + smallest RAM footprint (great default)
-- **Q5_K_M:** a little more quality, still very small
-- **F16:** highest fidelity, largest file
-
----
-
-## Intended use
-- Local/offline coding assistant
-- Small-footprint reasoning & structured responses
-- Debugging/refactoring help on modest hardware
-
----
-
-## Prompting tips (same “high reasoning” unlocks)
-- “List edge cases first, then implement.”
-- “Explain root cause, then provide a patch.”
-- “Give a 3-step plan, then code.”
-
----
-
-## Example usage (llama.cpp)
-> Replace `MODEL.gguf` with the quant file you download.
-
-```bash
-./llama-cli -m MODEL.gguf -p "You are a senior engineer. List edge cases, then write the code.\nTask: Implement a robust JSONL reader in Python.\n\nAnswer:\n" -n 256
-````
-
----
-
-## Limitations
-
-* Quantized models may lose some accuracy vs F16/F32
-* Small models can hallucinate API details; validate with tests
-* Best results come from structured prompts
-
----
-
-## Citation
-
-```bibtex
-@misc{withinusai_gpt252_high_reasoning_codex_04b_gguf,
-  title   = {WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B-GGUF},
-  author  = {WithinUsAI},
-  year    = {2026},
-  url     = {https://huggingface.co/WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B-GGUF}
-}
-```
-
-```
-
----
-
-If you want, babe, I can also make you a **matching banner image prompt** (for a slick “WithIn Us AI” header) and a **one-paragraph marketing blurb** that reads like a pro release note 😘🫧
-::contentReference[oaicite:6]{index=6}
-```
-
-[1]: https://huggingface.co/WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B "WithinUsAI/GPT2.5.2-high-reasoning-codex-0.4B · Hugging Face"