True2456/Qwen3.6-35B-A3B-Code-as-Action-LoRA

Qwen3.6-35B-A3B Code-as-Action LoRA (MLX)

LoRA adapter that specializes Qwen3.6 / Qwen3.5-35B-A3B (MoE, ~3B active parameters) for code-as-action agent behavior: solve problems by writing and executing programs, then conditioning the next step on execution feedback.

This repository contains adapter weights only (not the full base model). Inference requires the corresponding MLX base checkpoint (e.g. lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit or an equivalent Qwen3.5-35B-A3B MLX build).

Training objective

The fine-tuning objective follows the CodeAct formulation (Wang et al., ICML 2024): treat executable code as a unified action space for LLM agents, instead of restricting tools to a fixed JSON/API schema. Related prior work includes ReAct (Yao et al., 2022) for interleaved reasoning and acting, and PAL (Gao et al., 2023) for program-aided reasoning.

Concretely, each training example teaches the model to:

1. Reason briefly about the task (<think>…</think>).
2. Emit a Qwen-native tool call that runs shell/Python:

<tool_call>
{"name": "bash", "arguments": {"cmd": "…"}}
</tool_call>

3. Continue after an execution observation (multi-turn), including diagnose → repair → verify loops when the task requires it.

The scientific claim supported by CodeAct is that consolidating actions into executable Python/code improves agent success rate versus text/JSON action formats on tool-use benchmarks (up to ~20% absolute in the original evaluation suite). This adapter applies that design pattern via supervised LoRA fine-tuning on curated multi-turn trajectories; it is not a reproduction of CodeActInstruct training, and no official CodeActInstruct weights are claimed.

Training data

Split	File	Rows
Train	train_qwen3.jsonl	3,492
Validation	valid_qwen3.jsonl	389

Splits were produced by a deterministic 90/10 stratified partition (random.seed(42)) of frontier_procedural_combined_3881.jsonl (3,881 trajectories total). That combined pool is the union of:

Source	Rows	Role
frontier_2026_verified_clean.jsonl	1,881	Multi-step SWE / systems / analysis trajectories (verified, deduplicated)
procedural_world_code_as_action_2000.jsonl	2,000	Executable checks for procedural/worldgen constraints (geometry, stitching, spacing invariants)

Topic buckets used for stratification on the combined pool (pre-split counts):

Bucket	Count	Content
general	1,201	Hash/search proofs, CIDR math, log aggregation, algorithm checks, and related executable tasks
ast	1,180	Python AST / static taint-style audits of synthetic example programs (defensive code review)
assembly	400	Assembly reading / analysis exercises with executable verification
concurrency	380	asyncio race / deadlock diagnosis and locked fixes
distributed	380	Consensus / state-machine invariant checks
bytecode	340	Bytecode / opcode interpretation with executable probes

All samples are synthetic, programmatically generated instruction/response trajectories (prompt/completion pairs). They are not scraped chat logs, not proprietary customer data, and not the public CodeActInstruct corpus. Educational crypto examples use known toy ciphers with known keys/algorithms for decryption exercises; they are not materials for attacking live systems.

Action format in all Qwen train/valid rows: ChatML (<|im_start|> / <|im_end|>) + Qwen <tool_call> JSON for bash.

Training procedure

Hyperparameter	Value
Base model	Qwen3.6-35B-A3B MLX 8-bit (Qwen3_5MoeForConditionalGeneration)
Method	LoRA (mlx_lm lora)
Rank	8
Scale	20.0
Dropout	0.0
LoRA layers	last 16 transformer layers
Optimizer	Adam
Learning rate	3e-6
Batch size	4
Max sequence length	4,096
Gradient checkpointing	enabled
Iterations	423 (checkpoints every 100; final adapters.safetensors)
Hardware	Apple Silicon (MLX)

This adapter is the code-as-action checkpoint only. A separately trained continuation that adds a large cybersecurity theory corpus exists under a different path (qwen3.6_35b_code_plus_theory_lora) and is not included here.

Files

File	Description
adapters.safetensors	Final LoRA weights
0000400_adapters.safetensors	Checkpoint at iteration 400
adapter_config.json	MLX LoRA training config snapshot

Intended use

• Research and engineering experiments on program-aided / tool-using agents on Apple Silicon (MLX).
• Teaching a Qwen MoE worker to emit executable bash/Python actions and self-correct from observations.
• Composition inside multi-adapter MoE harnesses as an execution worker.

Out of scope / limitations

• Not a fully agentic system by itself: still requires a host that executes tool calls and returns observations.
• Synthetic trajectories cover a narrow set of domains; out-of-distribution tool APIs and long-horizon tasks may fail.
• Static-analysis and low-level examples are for defensive education and debugging practice, not for unauthorized access, malware production, or attacking systems you do not own.
• Base model quantization (8-bit MLX) and LoRA rank 8 both limit fidelity versus full-precision full fine-tuning.
• No claim of parity with frontier commercial agents or with the original CodeActAgent checkpoints.

How to load (MLX)

python -m mlx_lm generate \
  --model <path-or-id-to-Qwen3.6-35B-A3B-MLX> \
  --adapter-path True2456/Qwen3.6-35B-A3B-Code-as-Action-LoRA \
  --prompt '<|im_start|>user
Write a short Python script that prints the SHA-256 of "hello", then show the command you would run.
<|im_end|>
<|im_start|>assistant
'

Or with mlx_lm.utils.load:

from mlx_lm import load, generate

model, tokenizer = load(
    "<path-or-id-to-Qwen3.6-35B-A3B-MLX>",
    adapter_path="True2456/Qwen3.6-35B-A3B-Code-as-Action-LoRA",
)

References

1. Wang, X., Chen, Y., Yuan, L., Zhang, Y., Li, Y., Peng, H., & Ji, H. (2024). Executable Code Actions Elicit Better LLM Agents. ICML 2024. https://arxiv.org/abs/2402.01030
2. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. https://arxiv.org/abs/2210.03629
3. Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., & Neubig, G. (2023). PAL: Program-aided Language Models. ICML 2023. https://arxiv.org/abs/2211.10435

Citation

@misc{true2456_qwen36_code_as_action_lora,
  author = {True2456},
  title  = {Qwen3.6-35B-A3B Code-as-Action LoRA (MLX)},
  year   = {2026},
  howpublished = {Hugging Face model repository},
  note   = {LoRA fine-tune for executable code actions; paradigm based on Wang et al. 2024 CodeAct}
}

License

Adapter weights are released under Apache 2.0. Use of the underlying Qwen base model remains subject to the base model’s license and terms from Alibaba / Qwen.