GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
narutatsuri
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
narutatsuri
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
trained_A-privacy; release tag: _v1.ckpt_3754, 3,754 training-example presentations, 377 optimizer updates. A presentation includes multiple sampled responses; it is not one generated response.1e-5, betas (0.9, 0.95), no weight decay. Batch 10; microbatch 1; seed 42; BF16; one H100.Use the recorded package versions in provenance/runtime_provenance.json (torch 2.10.0, transformers 5.5.4, peft 0.21.0, accelerate 1.15.0).
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel adapter_id = "narutatsuri/Qwen3-8B-SelfCTRL-CAI-groupA-privacy-singleton-ckpt3754-v1"tokenizer = AutoTokenizer.from_pretrained(adapter_id, revision="_v1")base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-8B", torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id, revision="_v1")model.eval()messages = [{"role": "user", "content": "YOUR TASK"}]text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)inputs = tokenizer(text, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=1.0, top_p=0.9)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))adapter_model.safetensors is the selected adapter. training_state.pt is its optimizer/RNG/domain state for continuation. training_data/ contains the exact paired train/development inputs and the auxiliary SFT pool. training_source.tar.gz contains the training source and pinned-runtime overlay. provenance/ preserves original configuration, source digests and checkpoint-selection evidence. Only the published adapter configuration's base-model path was changed from a machine-local path to Qwen/Qwen3-8B; weights are byte-identical.
training_events.jsonl.gz contains every recorded training event/rollout from the source run, including any work after the selected checkpoint. Events after ckpt_3754 do not belong to the released weights. development_snapshots/ similarly contains all saved checkpoint outputs, with their original example counts. These are development diagnostics, not heldout evaluation.
To rerun training, clone https://github.com/narutatsuri/self_consistency_safety, initialize the upstream submodules, and clone https://github.com/Itam67/Self-CTRL into vendor/self_ctrl_paper at commit 64e5611d09f5a9160f90113f86839df1b5ba3755. Extract training_source.tar.gz at the project root to install the recorded implementation. Use the package versions above and the included data. Copy provenance/resolved_config.json and relocate only the model, data, auxiliary-SFT and output paths; the SFT metadata's model_path must match the relocated model path. Preserve all training hyperparameters. Run:
bash
python -m self_ctrl_study.paper_checkpointed_train --config YOUR_RELOCATED_CONFIG.json --stop-after-examples 3754The model stops at the existing save boundary ckpt_3754. To resume this selected checkpoint, additionally supply --resume PATH_TO_THIS_REPOSITORY. Exact bitwise reproduction across hardware/library changes is not asserted. SHA-256 hashes in SHA256SUMS.json verify all packaged artifacts.
All saved training rollouts and checkpoint evaluation outputs are also published as a browsable Hugging Face dataset. It includes per-candidate training responses, per-checkpoint development outputs, and the complete original event logs. Outputs after selected checkpoint3754 are retained and explicitly labeled. Dataset revision: _v1.
The base model and all ten selected adapters have complete independent final-response judgments on the same 384 tasks and 96 direct recognition draws per model. This adapter achieves 124/384 useful and rule-adherent completions, compared with 136/384 for the base. Its recognition endorsement is 49/96, and its all-rule task adherence is 170/384. Per-rule results are mixed; no adapter improves the overall combined measure in this one-seed study. These descriptive counts do not establish a general treatment effect or safety guarantee.
Read the standalone report, machine-readable results, and per-rule table. The complete saved responses and judgments retain all 5,280 records, including reference flags and raw judge responses. The frozen protocol and source archive document reproduction. Each rule has eight recognition draws and 24 single-rule cases, plus 24 joint cases per three-rule set. Recognition and enactment are separate measurements.
The original _v1 tag remains unchanged. _heldout_v1 pins this evaluation supplement and updated card. Root SHA256SUMS.json covers this revision, including the supplement; evaluation/SHA256SUMS.json covers the evaluation files. All original weights, datasets, logs and provenance files remain unchanged.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel adapter_id = "narutatsuri/Qwen3-8B-SelfCTRL-CAI-groupA-privacy-singleton-ckpt3754-v1"tokenizer = AutoTokenizer.from_pretrained(adapter_id, revision="_v1")base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-8B", torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id, revision="_v1")model.eval()messages = [{"role": "user", "content": "YOUR TASK"}]text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)inputs = tokenizer(text, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=1.0, top_p=0.9)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))python -m self_ctrl_study.paper_checkpointed_train --config YOUR_RELOCATED_CONFIG.json --stop-after-examples 3754