Not-For-All-Audiences

Mistral-7B-Instruct-v0.3 · Artisan

A LoRA fine-tune of mistralai/Mistral-7B-Instruct-v0.3 tuned for direct, practical assistant responses with a side of grounded character roleplay. Built on a curated seed of hand-written technical Q&A and dialogue, then balanced against general-purpose and conversational sets so the model keeps its broad capabilities.

What it's good at

Technical answers without the boilerplate. Backend / systems / Go / databases / networking — concise, no "as a language model…" preamble.
In-character dialogue. Roleplay scenes with continuity, restraint, and tone awareness. Won't break frame.
Adult-adjacent content. Open by default; trained with a curated explicit-RP slice. Not for all audiences.

MLX variants

Folder	Bits	Size	Quality	Load
`mlx-4bit-dwq/`	4	~4.1 GB	best at 4-bit	`mlx_lm.generate --model darthcrawl/mistral-7b-instruct-v0.3-artisan/mlx-4bit-dwq`
`mlx-4bit/`	4	~4.1 GB	vanilla 4-bit	`mlx_lm.generate --model darthcrawl/mistral-7b-instruct-v0.3-artisan/mlx-4bit`
`mlx-6bit/`	6	~6.0 GB	near-lossless	`mlx_lm.generate --model darthcrawl/mistral-7b-instruct-v0.3-artisan/mlx-6bit`
`mlx-8bit/`	8	~7.5 GB	effectively FP16	`mlx_lm.generate --model darthcrawl/mistral-7b-instruct-v0.3-artisan/mlx-8bit`

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "darthcrawl/mistral-7b-instruct-v0.3-artisan"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [
    {"role": "user", "content": "Explain consistent hashing in two paragraphs."},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
out = model.generate(inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

MLX quants (Apple Silicon)

Quantized variants for mlx-lm:

4-bit, 4-bit DWQ, 6-bit, 8-bit

Training

Category	Notes
Method	QLoRA — 4-bit base + LoRA r=16, alpha=32, dropout=0.05
Targets	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Data mix	~3% curated artisan, ~14% character RP, ~83% general-purpose
Schedule	2 epochs, cosine, lr 2e-4, warmup 50, effective batch 16
Hardware	Single H100 80GB

Limitations

English only.
Inherits the base model's biases and knowledge cutoff.
Adult RP slice means the model is more permissive than the stock instruct. Use accordingly.

License

Apache 2.0, inherited from the base model.