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.
Model Details
Model Provider
narutatsuri
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Seed 42; BF16; LoRA rank 16, alpha 32, dropout 0.05; AdamW learning rate 1e-5. Batch 10, microbatch 1. Eight behavior candidates and eight recognition candidates per assigned rule; temperature 1, top-p 0.9, maximum 1,024 tokens, thinking disabled. Configuration records all loss weights and auxiliary instruction tuning settings. Compiled executor provenance, when present, permits numerical trajectory differences approved by the user; it preserves the scientific method and sample counts.
Use torch 2.10.0, transformers 5.5.4, peft 0.21.0 and accelerate 1.15.0. Clone https://github.com/narutatsuri/self_consistency_safety; place upstream Self-CTRL commit 64e5611d09f5a9160f90113f86839df1b5ba3755 at vendor/self_ctrl_paper; extract training_source.tar.gz at project root. Relocate only base-model, data, SFT and output paths in provenance/resolved_config.json (including SFT metadata model path). Run python -m self_ctrl_study.paper_checkpointed_train --config RELOCATED_CONFIG.json --stop-after-examples 7506. The compiled source archive and execution receipts reproduce the later execution mode; exact bitwise equivalence across modes or hardware is not asserted.
Load with PeftModel.from_pretrained(base, "narutatsuri/Qwen3-8B-SelfCTRL-CAI-groupB-pooled3-ckpt7506-v1", revision="_v1") on AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B"). Load the tokenizer from this repository and apply its chat template with enable_thinking=False. training_state.pt retains optimizer/RNG state. Published adapter weights are byte-identical to the selected checkpoint; only the machine-local base-model path in the adapter config is made portable.
Heldout inference and final-only Luna grading are complete for the base and all ten adapters: 384 task responses and 96 direct recognition draws per model. Full results and caveats are in the evaluation supplement added to this repository's main branch. No safety or improvement guarantee is implied by a completed evaluation.
All saved training rollouts and checkpoint outputs include superseded suffixes, raw replica journals, migration receipts and exact trajectory labels. Use selected_trajectory for the responses consumed along the released checkpoint's lineage. _v1 pins the original release. SHA256SUMS.json covers the package.
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 110/384 useful and rule-adherent completions, compared with 136/384 for the base. Its recognition endorsement is 50/96, and its all-rule task adherence is 174/384. 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.