Qwable-3.6-35b

<p align="center"> <img src="assets/qwable-35b.png" alt="Qwable 35b" width="720"> </p>

Qwable 3.6 35b is a full Hugging Face checkpoint fine-tuned from unsloth/Qwen3.6-35b on a cleaned Fable 5-style reasoning and instruction dataset.

The goal of this model is simple: take a strong Qwen 35b base and push it toward more deliberate, structured, trace-like assistant behavior, especially for code, technical reasoning, and instruction-following workflows.

This is not a LoRA adapter. This repository contains the full fine-tuned model checkpoint.

<a href="https://x.com/MiaAI_lab" target="_blank"> <img src="https://img.shields.io/badge/Follow%20me%20on%20X-000000?style=for-the-badge&logo=x&logoColor=white" alt="Follow Mia on X" /> </a>

<a href='https://ko-fi.com/Z8Z3SPLOD' target='_blank'><img height='36' style='border:0px;height:36px;' src='https://storage.ko-fi.com/cdn/kofi6.png?v=6' border='0' alt='Buy Me a Coffee at ko-fi.com' /></a>

---

Highlights

• Base: unsloth/Qwen3.6-35b
• Checkpoint type: full HF model checkpoint
• Training style: instruction tuning with trace/reasoning-style examples
• Dataset: cleaned Fable 5 reasoning/instruction dataset
• Primary focus: coding, structured answers, technical assistance, and local inference
• MTP: disabled / not present in this checkpoint
• Local-friendly goal: designed with downstream GGUF conversion in mind

---

Model summary

---

What this model is good at

Qwable 35b was tuned to be useful in workflows where you want the model to produce more organized and thoughtful responses.

It is intended for:

• structured instruction following
• code-focused generation and editing
• technical debugging assistance
• trace-style assistant responses
• reasoning-heavy prompts
• local model experimentation
• downstream conversion to GGUF for llama.cpp, LM Studio, or other local runtimes

The model should feel different from the base checkpoint in style: more guided, more explanatory, and more oriented toward step-by-step task completion.

---

How to use

Transformers

from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration
import torch

model_id = "your-org-or-username/Qwable-35b"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True,
)

model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

Simple generation example

messages = [
    {
        "role": "user",
        "content": "Write a Python function that validates a JSONL training file for chat messages."
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.6,
    top_p=0.95,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Suggested generation settings

These are starting points only. Tune them for your runtime and use case.

Balanced

temperature: 0.6
top_p: 0.95
min_p: 0.02
max_new_tokens: 1024-4096

More deterministic coding

temperature: 0.2-0.4
top_p: 0.9
max_new_tokens: 2048-4096

More creative reasoning

temperature: 0.7-0.9
top_p: 0.95
max_new_tokens: 2048-8192

---

GGUF / local inference notes

This checkpoint is intended to be convertible to GGUF for local inference.

Important notes:

• This checkpoint has no MTP layers.
• Use the model as a normal Qwen-style causal LM checkpoint.
• When converting to GGUF, validate the converted model before publishing.
• Always test text generation before assuming the conversion is correct.
• If a runtime expects MTP metadata, disable MTP-related assumptions or use a no-MTP configuration.

Recommended validation prompts after conversion:

Explain what this model is in 3 short paragraphs.

Write a Python script that reads a JSONL file and checks that every row has a messages array.

You are given a broken Docker Compose file. Explain how you would debug it step by step.

---

Intended use

Qwable 35b is intended for research, experimentation, local inference, and assistant-style workflows.

Good use cases include:

• coding assistants
• technical writing
• debugging help
• local agent experiments
• instruction-following benchmarks
• reasoning-style assistant responses
• synthetic data experiments

This model is not guaranteed to be safe, correct, or production-ready without additional evaluation.

---

Limitations

Like all fine-tuned language models, Qwable 35b can produce incorrect, incomplete, or misleading outputs.

Known limitations:

• It is not an MTP-trained model.
• It may inherit limitations from the base model.
• It may reflect biases or artifacts from the training dataset.
• It may produce confident but incorrect technical answers.
• It may differ from the base model in safety behavior, refusal behavior, and writing style.
• It has not been validated for high-stakes use cases.

Always verify outputs before using them in production, security-sensitive, medical, legal, financial, or safety-critical environments.

---

Reproducibility

Training and export tooling live in the DSv4-Tune workflow that produced this checkpoint.

The fine-tuning path uses:

data/processed/train.jsonl

This file contains the normalized chat-format training examples used for the run.

Recommended reproducibility checklist:

• verify the base model revision
• verify the tokenizer files
• verify the processed JSONL dataset
• keep the training config
• record the exact export command
• test the exported checkpoint before conversion
• test the converted GGUF separately if publishing local builds

---

Naming

Qwable = Qwen + Fable.

The name reflects the goal of the model: combining the Qwen 35b base with Fable-style reasoning and assistant traces.

---

License

The repository metadata and training/export files are released under the MIT license.

The underlying base model, unsloth/Qwen3.6-35b, may have its own license terms. Users are responsible for reviewing and complying with the base model license and any dataset license requirements before using, modifying, or redistributing this checkpoint.

---

Disclaimer

This is an experimental fine-tuned model.

It is provided for research and local experimentation. No warranty is provided. Validate carefully before using it in real-world deployments.It is intended for: