LLM-OS-Models/Qwen3.5-9B-Graph-Preflexor-ORPO

Merged full model from ORPO cold-start stage of the Graph-PRefLexOR reproduction fork gyunggyng/lfm-graph-preflexor (fork of lamm-mit/graph-preflexor-grpo, arXiv 2607.00924v1).

• Base model: principled-intelligence/Qwen3.5-9B-text-only (Qwen3_5TextForCausalLM, model_type: qwen3_5_text, hybrid linear + full attention)
• Stage: ORPO cold-start (step 1 of 2; Graph-GRPO refinement pending)
• Architecture: text-only Qwen3.5 — 32 layers, 4 full-attention layers (every 4th), 28 linear-attention layers, hidden 4096, vocab 248087.
• Checkpoint: checkpoint-250 merged into the base.

Training

• Framework: TRL 0.24 ORPOTrainer, PEFT LoRA (r=32, alpha=64, dropout=0.05, targets = q/k/v/o/gate/up/down).
• Data: lamm-mit/graph_reasoning_10K filtered for graph-reasoning items with structured <brainstorm>...<synthesis> reasoning targets.
• Hardware: 4× H200 (CUDA 12.8), torch 2.12.0.dev20260407+cu128, transformers 5.5.4, bfloat16.
• Hparams: LR 5e-6, effective batch 8 (per_device 1 × accum 2 × world 4), max_prompt 1536, max_completion 4096, eval disabled to fit VRAM (ORPO's concatenated_forward OOMs at 9B + seq 5632).

Results (checkpoint-250)

Eval was run with scripts/05c_eval_transformers.py (4-GPU shard-parallel transformers, eager attention, thinking enabled) because vLLM 0.19 / 0.20 do not register Qwen3_5TextForCausalLM — see "Known limitations" below.

Output format

The model emits a structured reasoning trace inside <think>:

<think>
<brainstorm>... free-form exploration ...</brainstorm>
<graph>... concept graph (natural language) ...</graph>
<graph_json>{"nodes": [...], "edges": [...]}</graph_json>
<patterns>... reusable abstractions ...</patterns>
<synthesis>... integrated reasoning ...</synthesis>
</think>
final answer

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "LLM-OS-Models/Qwen3.5-9B-Graph-Preflexor-ORPO",
    dtype="bfloat16", device_map="auto", trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained(
    "LLM-OS-Models/Qwen3.5-9B-Graph-Preflexor-ORPO",
    trust_remote_code=True,
)
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Your graph-reasoning question here"}],
    tokenize=False, add_generation_prompt=True,
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=3500, do_sample=True, temperature=0.2)
print(tok.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=False))

attn_implementation="eager" is required if flash-linear-attention is not installed; SDPA silently returns empty tokens otherwise.

Known limitations

• vLLM: Qwen3_5TextForCausalLM is not in vLLM's registered architectures as of vLLM 0.20.2 (only Qwen3_5ForConditionalGeneration / Qwen3_5MoeForConditionalGeneration / Qwen3_5MTP are). vLLM's generic TransformersForCausalLM wrapper also fails because it expects the multimodal prefix model.language_model.*, while text-only weights are flat at model.layers.*. Use transformers for inference until vLLM adds native text-only Qwen3.5 support.
• Final eval OOM: ORPOTrainer forces a final evaluate() after training, which OOMs on 9B + seq 5632 in concatenated_forward. Checkpoint-250 (saved before the crash) is what's merged here.
• GRPO stage: not yet completed. The Qwen3.5 text-only arch blocks the vLLM-based rollout server the GRPO config assumes, so the GRPO refinement run was abandoned at the import step.

Citation

@article{graphpreflexor2025,
  title={Graph-PRefLexOR: Graph-based Preference-based Reasoning via Learning},
  doi={10.48550/arXiv.2607.00924}
}