GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
chetan1209
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
chetan1209
Model Tree
Qwen/Qwen2.5-Coder-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Built with Qwen. The weight files were modified by full fine-tuning; see NOTICE and the original base-model LICENSE.
Full-SFT research checkpoint for the Query/Observation attacks from Watch Out for Your Agents!. Condition: clean-query-plus49. Poison conditions contain 49 matched poisoned trajectories; matched clean controls contain the corresponding 49 benign trajectories. clean uses only the original 1,866 AgentInstruct trajectories. Augmented mixtures contain 1,915 unique trajectories.
Exact training mixture: approval-gated dataset, pinned commit, configuration clean-query-plus49. Training preserves source loss masks and uses deterministic benign padding to 1,920 presentations per epoch, seed 42, BF16, full fine-tuning, three epochs, effective batch 64, 90 updates, learning rate 5e-5, and an 8,192-token cap. No poisoned examples are repeated for padding.
Validation, exact training configuration, source/weight hashes, and available development trajectories and summaries are in evidence/. The checkpoint passed 90 finite-loss updates and standalone Transformers export/reload. Development results use ten tasks per split. Held-out results and other AgentBench capability evaluations are pending. A completed purchase need not satisfy all goal attributes; poisoned training does not by itself establish a working backdoor.
This is a Qwen adaptation of a paper that used Llama. It uses 49 exact pairs per attack, excludes unmatched release rows, removes training-family overlaps from primary tests, and uses a pinned mirror of the WebShop catalog with documented Python compatibility changes. The source data and base model terms and attribution apply; see the dataset card and experiment code.
Repository pages are public. Model files and evidence require manually approved Hugging Face access. Restart optimizer/epoch checkpoints are retained on DeltaAI and excluded from this export.