cfontes/GLM-5.2-F5-Molt-LoRA

GLM-5.2-F5-Molt — LoRA Adapters 🛡️

LoRA adapters for the GLM-5.2-F5-Molt cybersecurity fine-tune. These let you serve the F5-Molt behavior on top of a stock zai-org/GLM-5.2 (FP8) base without downloading the full 698 GB merged checkpoint.

If you just want the ready-to-run model, grab the merged version: cfontes/GLM-5.2-F5-Molt.

---

Adapters in this repo

Subdir	Stage	Rank	Alpha	Beta	Notes
sft/	SFT	64	128	—	Multi-teacher distillation (reasoning + cyber Q&A)
dpo-v1/	DPO v1	16	32	0.1	First preference-alignment pass
dpo-v2/	DPO v2	16	32	0.1	Refined preference pairs, init_from=sft
combined-v2/	SFT + DPO-v2 (merged)	—	—	—	Recommended — single adapter, best results

Use combined-v2/ for the best results. It folds the SFT reasoning gains and the DPO-v2 refusal-boundary alignment into one adapter you can load in a single step.

All adapters target attention projections (q/kv a/b projections and o_proj) on the top 18 layers (layers 60–77) of GLM-5.2. The base path recorded in the adapter configs is the training scratch path; point base_model_name_or_path at zai-org/GLM-5.2 (or your local FP8 copy) when loading.

---

Configuration details

SFT (sft/)

• r=64, lora_alpha=128, lora_dropout=0.0
• 160 steps · lr 2e-4 · seq_len 2048
• Target modules: attention projections, layers ≥ 60 (90 modules)

DPO v1 (dpo-v1/)

• r=16, lora_alpha=32, beta=0.1
• lr 5e-5
• Preference alignment on legitimate-but-refused security requests

DPO v2 (dpo-v2/)

• r=16, lora_alpha=32, beta=0.1
• 30 steps · lr 5e-5 · init_from=sft (SFT adapter frozen as the DPO reference policy)
• Trained on 897 curated preference pairs (chosen = helpful security answer, rejected = refusal/hedge)

combined-v2 (combined-v2/)

• SFT + DPO-v2 folded into one adapter for single-step loading (~234 MB)

---

Loading

PEFT (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "zai-org/GLM-5.2"
tok = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype="auto", device_map="auto", trust_remote_code=True
)

# Recommended: the combined adapter
model = PeftModel.from_pretrained(base, "cfontes/GLM-5.2-F5-Molt-LoRA", subfolder="combined-v2")
model.eval()

vLLM

vllm serve zai-org/GLM-5.2 \
  --enable-lora \
  --lora-modules f5molt=cfontes/GLM-5.2-F5-Molt-LoRA/combined-v2 \
  --max-lora-rank 64 \
  --trust-remote-code

Then request with "model": "f5molt". (Bump --max-lora-rank to 64 to accommodate the SFT adapter; the DPO adapters are rank 16.)

SGLang

python -m sglang.launch_server \
  --model-path zai-org/GLM-5.2 \
  --lora-paths f5molt=cfontes/GLM-5.2-F5-Molt-LoRA/combined-v2 \
  --max-lora-rank 64 \
  --trust-remote-code

Reference the adapter by name (f5molt) in your requests. To A/B individual stages, add more --lora-paths entries (e.g. sft=.../sft dpo2=.../dpo-v2).

Rank note: if you load the standalone sft/ adapter, serving engines need --max-lora-rank 64. The DPO adapters and combined-v2 fit within rank 16/64 respectively.

---

Refusal boundary

These adapters shift GLM-5.2 to answer legitimate security work (vulnerability research, malware analysis, reverse engineering, exploit/tool development). The alignment only refuses self-harm instructions and child sexual exploitation content. Use responsibly and legally.

License

GLM license, inherited from zai-org/GLM-5.2. See the license link.

Built on zai-org/GLM-5.2. Distilled from Claude Opus 4.7/4.8 reasoning traces and the Fable 5 corpus. Aligned with Molt.