stabilizer-mini-v1

stabilizer-mini-v1

Overview

This is a test model.

Technical notes

• Base: openai/gpt-oss-20b (bf16)
• Steering: rank-1 delta on Q/K/V across 24 layers (RMSNorm-aware)
• Concept vector: concept_vec_v15k.pt, shape [24, 6, 2880], gain=0.5
• Checkpoint: single baked weights (no LoRA/adapters; knowledge ≈ base)
• Data used: neutral_examples=86376, pairs_used=14400
• Source files: narukijima/stabilizer → S_instruction_pairs_en.jsonl, S_instruction_pairs_ja.jsonl
• Inference: use base tokenizer & chat template

Quick inference

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
M = "narukijima/stabilizer-mini-v1"
tok = AutoTokenizer.from_pretrained(M, trust_remote_code=True)
mdl = AutoModelForCausalLM.from_pretrained(
    M, torch_dtype=torch.bfloat16, device_map='auto', trust_remote_code=True
)
msgs = [{"role":"user","content":"test"}]
p = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = mdl.generate(**tok(p, return_tensors='pt').to(mdl.device),
                   max_new_tokens=64, do_sample=True, temperature=0.7)
print(tok.decode(out[0], skip_special_tokens=True))