---
language:
- en
license: apache-2.0
tags:
- text-generation
- pytorch
- causal-lm
- reasoning
- coding
- glm-distillation
- mature
base_model: distilgpt2
---

# distilgpt2-glm5-reasoning-coder

## Model Details

### Model Description
This model is a full fine-tuned version of **DistilGPT2**, exposed to an aggressive curriculum of GLM-4 and GLM-5 level distillation traces, high-reasoning frameworks, comprehensive coding logic, and mature internet discourse. It is designed to act as a highly responsive, analytical engine capable of deep structural reasoning and complex logic emulation.

Trained natively at an accelerated maximum learning rate with a cosine decay schedule, the model synthesizes diverse programmatic and theoretical domains from a massive multi-repository corpus, processed at the model's absolute maximum context window of 1024 tokens.

- **Developed by:** GODsStrongestSoldier
- **Model type:** Causal Language Model (Transformer Decoder)
- **Language:** English
- **License:** Apache 2.0
- **Finetuned from model:** `distilgpt2`

---

## Datasets Used for Fine-Tuning
This model was trained comprehensively on a synthesized Golden Mix dynamically extracted from the following datasets:
- [bmeyer2025/glm5-reasoning-traces](https://huggingface.co/datasets/bmeyer2025/glm5-reasoning-traces)
- [WithinUsAI/gods_universe_codex_distill_god_seed_25k](https://huggingface.co/datasets/WithinUsAI/gods_universe_codex_distill_god_seed_25k)
- [WithinUsAI/GOD_Coder_Complete_DataSet](https://huggingface.co/datasets/WithinUsAI/GOD_Coder_Complete_DataSet)
- [jjmachan/NSFW-reddit](https://huggingface.co/datasets/jjmachan/NSFW-reddit)

---

## Training Details

### Training Procedure
The model underwent **full fine-tuning** without the use of adapters or LoRA layers. All native parameters of the base model were globally updated. The training harness dynamically parsed heavily nested dataset repositories, enforcing a strict shape constraint to generate mathematically perfect 1024-token continuous sequences for the GPU, maxing out the DistilGPT2 context window.

To maximize adaptation to the GLM-level reasoning data, an absolute peak learning rate (`3e-4`) was utilized alongside a 5% warmup phase and a cosine scheduler. Local checkpoints were strictly enforced to optimize Kaggle disk performance.

#### Hardware
- **Environment:** Kaggle
- **Accelerators:** Dual NVIDIA T4 GPUs (15GB VRAM each)

#### Hyperparameters
- **Epochs:** 1 
- **Context Window / Block Size:** 1024
- **Per-Device Batch Size:** 4
- **Gradient Accumulation Steps:** 16
- **Effective Global Batch Size:** 128
- **Peak Learning Rate:** 3e-04
- **Learning Rate Scheduler:** Cosine
- **Warmup Ratio:** 0.05
- **Optimizer:** Fused AdamW (`adamw_torch_fused`)
- **Mixed Precision:** fp16
- **Gradient Checkpointing:** Enabled