ElijahSdjsiaofjdasiofjs/chairreasoner-full-sft-gemma

ChairReasoner — Full SFT on Gemma 4 E4B (MLX)

A full-parameter supervised fine-tune of Google Gemma 4 E4B on a diverse synthetic reasoning and instruction dataset, converted to MLX for Apple Silicon.

Model: ElijahSdjsiaofjdasiofjs/chairreasoner-full-sft-gemma
Base: google/gemma-4-E4B (4.5B effective / ~8B total params)
Format: MLX, bfloat16 (unquantized) — optimised for Apple Silicon Macs
Context length: 128K tokens
License: Apache 2.0 (inherited from Gemma 4)

---

Model Details

Why full SFT?

Unlike LoRA or QLoRA adapters which only train a small subset of parameters, this run performs true full-parameter SFT — every weight in the 8B-parameter model is updated. This was done on a 16 GB Apple Silicon Mac using MLX with aggressive memory management (lazy weight loading, gradient checkpointing, short sequence lengths).

Architecture

Property	Value
Base model	Gemma 4 E4B (dense)
Total parameters	~8B (4.5B effective)
Layers	42 (hybrid sliding + full attention)
Hidden size	2560
Attention heads	8 × 256-dim
KV heads	2 (with 18 KV shared layers)
Sliding window	512
Context length	128K
Vocabulary	262K tokens (SentencePiece)
Precision	bfloat16

Training Configuration

Hyperparameter	Value
Optimizer	AdamW
Learning rate	2e-5
Batch size	1 (gradient accumulation: 8)
Max sequence length	512
Total iterations	1500
Hardware	Apple Silicon Mac (16 GB unified memory)
Framework	MLX + mlx-lm

Inference Settings (recommended)

temperature = 1.0
top_p = 0.95
top_k = 64

---

Training Data

The model was fine-tuned on a synthetic multi-turn instruction dataset covering a broad range of expert domains. Each conversation is grounded in a specific system persona.

Dataset stats:

Stat	Value
Training examples	8,706
Validation examples	435
Multi-turn conversations (3+ messages)	~95%
Max conversation turns	13
Average messages per conversation	4.1
Total message turns	35,373
Unique system personas	5,516

Domain Coverage

The data spans a wide range of expert domains, each with tailored system prompts:

• 🧪 Science & Technology — physics, chemistry, biology, engineering
• 💻 Coding & Engineering — software engineering, DevOps, data engineering, frontend, security
• 📚 Humanities — philosophy, history, comparative literature, political science
• ✍️ Creative Writing — narrative craft, comedy, unreliable narrators, story structure
• 🧠 Cognitive Science — psychology, behavioral ecology, neuroscience
• 🏥 Medicine — oncology, clinical decision-making, pharmacology
• 💰 Economics & Social Science — economic theory, game theory, research methodology
• 🤖 AI & ML — machine learning, reasoning, agentic systems

Thinking Mode

Training data leverages Gemma 4's built-in reasoning capability via <think> tags, where the model produces step-by-step reasoning before answering:

<think>
Step-by-step reasoning...
</think>
Final response.

---

Usage

The model is stored in MLX safetensors format and is intended for use with MLX and mlx-lm.

Loading with MLX

from mlx_lm.utils import load_model, load_tokenizer

model, _ = load_model("/path/to/chairreasoner-full-sft-gemma")
tokenizer = load_tokenizer("/path/to/chairreasoner-full-sft-gemma")

Loading with Transformers (via conversion)

The model originates from the Transformers-compatible Gemma 4 format. To use it with Hugging Face Transformers, convert the MLX weights back or download the base separately.

---

Technical Notes

• This model was trained fully offline — no internet access during training (HF_HUB_OFFLINE=1)
• Weights are unquantized bf16 — the model.safetensors file is ~15 GB
• The training script (train.py) includes full-parameter unfreezing via model.unfreeze(), as opposed to the default LoRA mode in mlx-lm
• Training was tuned to fit within 16 GB unified memory on Apple Silicon using memory-mapped weight loading and aggressive Metal cache clearing
• Checkpoints (if any) are stored in checkpoints/ as full weight dumps

---

Repository Structure

├── README.md                          # This file
├── train.py                           # Full SFT training script
├── mlx_unquantized_gemma/             # Base model + fine-tuned weights
│   ├── model.safetensors              # ~15 GB unquantized weights
│   ├── config.json                    # Model configuration
│   ├── generation_config.json         # Generation hyperparameters
│   ├── tokenizer.json                 # SentencePiece tokenizer
│   ├── tokenizer_config.json          # Tokenizer configuration
│   ├── processor_config.json          # Multimodal processor config
│   ├── chat_template.jinja            # Chat template (with thinking support)
│   └── .gitattributes                # Git LFS config
├── local_dataset/                     # Original Hugging Face Arrow dataset
│   ├── data-00000-of-00001.arrow      # 66 MB — source training data
│   ├── dataset_info.json              # Dataset metadata
│   └── state.json                     # Dataset state
├── mlx_data/                          # Cached JSONL files
│   ├── train.jsonl                    # 64 MB — 8,271 training examples
│   └── valid.jsonl                    # 3.4 MB — 435 validation examples
├── checkpoints/                       # Training checkpoints (if resumed)
└── model_params/                      # Reserved for parameter exports

---

Acknowledgements

• Google DeepMind for the Gemma 4 model family (Apache 2.0)
• Apple MLX team for the MLX framework enabling full SFT on Mac
• Built with mlx-lm

---

License

This repository is licensed under Apache 2.0, inherited from the Gemma 4 base model license.