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-spark-step2-a2-step2-r1-c6-rl
This is the full BF16 checkpoint after step2 recovery round 1, cycle 6 DPO and 473 response-level RL updates. It was exported on 2026-09-26 UTC.
- Original campaign starting point:
hayabusa-9b-temporal-v2-step1-r4-c8-rl - Immediate RL base:
hayabusa-9b-spark-step2-a2-step2-r1-c6(cycle 6 DPO) - RL batch: 10 groups, 195 rollouts, 473 scored response updates
- Training context ceiling: 32K tokens; six overlong prompts were trimmed from the start to retain their complete responses
- Serving context: 64K tokens with prompt caching
- Format: full-weight BF16 safetensors, 32 shards and 427 model tensors
- Saved architecture:
Qwen3_5ForCausalLM/qwen3_5_text - Transformers version recorded by the checkpoint:
5.3.0.dev0
For reproducible deployment, use the immutable Hub revision hayabusa-9b-spark-step2-a2-step2-r1-c6-rl.
The SHA256SUMS and checkpoint_provenance.json files record the export's
integrity and training provenance. The native RL export contains model and
tokenizer files, without optimizer resume state.
Student Evaluation
The step2 student probe passed in 93 of 200 debugger steps. Public, private,
and regression checks passed. This result is from the full debugging agent;
Hayabusa selected debugger actions while director, expert, and developer roles
used openai/gpt-6-luna. Step3 evaluation was still in progress when this
revision was prepared. This is 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.