hayabusa-9b

Hayabusa 9B

<p align="center"> <img src="assets/hayabusa9b.png" alt="Hayabusa 9B logo" width="260"> </p>

Hayabusa 9B is a text-only, full-weight fine-tune of Qwen/Qwen3.5-9B for software debugging and structured agent action selection.

This repository contains a full merged BF16 Transformers checkpoint, not a LoRA, QLoRA, or adapter. It can be loaded directly by Transformers-compatible runtimes or used as the source checkpoint for MLX conversion.

Current Checkpoint

hayabusa-9b-spark-pre-stochastic-restore-step3-r4-c6-online-sft-judge-v1-reboot-20261022a

Completed step3 recovery round 4, Cycle 6 final SFT, exported on 2026-10-27 UTC. Checkpoint identity: 34142531dc47fdad8e88a3585392c25dcd67fd64566dee96c2806fd7d81057fa.

• Cycle 6 training: 155 DPO updates, then 289 RL updates, then 55 original teacher SFT updates.
• These are the exact weights used by the latest student run: step4 initial attempt 1. No step4 training is included.
• Full-weight BF16 safetensors: 32 shards, 427 model tensors. Stored model weights are exported without folding optimizer residuals.
• Training context ceiling: 32K tokens with left prompt truncation. Loss covers only the final assistant reasoning and content; earlier messages are masked.
• Saved architecture: Qwen3_5ForCausalLM / qwen3_5_text; recorded Transformers version 5.3.0.dev0.
• Weights and tokenizer only: no optimizer, scheduler, RNG state, private training examples or execution logs are published.

For a stable revision, use step3-r4-c6-final-sft-34142531. SHA256SUMS and checkpoint_provenance.json describe the published files and completed training lineage. Successful export and training audits are integrity checks, not behavioral-quality or independent-mastery claims. The step3 final student attempt and step4 initial student attempt did not pass within their 200-step budgets; neither test is claimed mastered.

Intended Use

Hayabusa is specialized for structured debugger contexts containing source, logs, traces, test outcomes, compact progress history, and constrained actions. It is not intended as a general-purpose chat model.

The checkpoint is suitable for download and conversion by MLX tooling because all model weights, tokenizer files, chat template, generation configuration, and safetensors index are present at repository root.

Loading With Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "georvn7/hayabusa-9b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Select the next grounded debugging action."}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Limitations

• Specialized debugging/action model, not a broad assistant benchmark release.
• Best results require the structured runtime context and action schema used during training.
• The model remains experimental for long-horizon autonomous debugging.
• Vision inputs are unsupported.

License

This model inherits the upstream Qwen/Qwen3.5-9B Apache-2.0 license. See LICENSE.