georvn7/hayabusa-9b

Hayabusa 9B

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-temporal-v2-step1-r4-c8-rl

Final checkpoint from Step 1, recovery Round 4, Cycle 8, including Cycle 8 DPO and the subsequent checkpointed RL update. The cycle used the Round 0 clean anchor.

• Artifact completed: 2026-09-12 UTC
• DPO training base: hayabusa-9b-temporal-v2-step1-r4-c7-rl
• RL training base: hayabusa-9b-temporal-v2-step1-r4-c8 (Cycle 8 DPO)
• DPO rows: 60; policy-branch tokens: 1,343,631
• RL rows: 39; policy sequences: 89; optimizer steps: 89/89
• RL training tokens: 1,307,203; truncated sequences: 0
• Training context ceiling: 32K tokens; evaluation serving context: 64K tokens
• Format: full-weight BF16 safetensors, 32 shards
• Assistant thinking included in completion training; prompt loss masked
• Saved architecture: Qwen3_5ForCausalLM / qwen3_5_text
• Transformers version recorded by the checkpoint: 5.3.0.dev0
• Trainer source commit: d154df4a9b2372de7f4b477e06a256e4c0292688

For reproducible deployment, use the Hub revision hayabusa-9b-temporal-v2-step1-r4-c8-rl instead of mutable main. File checksums are recorded in SHA256SUMS and checkpoint_provenance.json.

Final Student Evaluation

The final Step1 evaluation did not pass after 200 total debugger steps (33 before interruption plus 167 after resumption). All 63 test runs failed; regression tests were not run because the target did not pass.

This checkpoint selected debugger actions. Supporting director, expert, and developer roles used openai/gpt-5.6-luna. The debugger output budget was 4,096 tokens. This is a specialized runtime evaluation, not a standalone coding benchmark. Training data, execution logs, and private project source are not included.

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.