bridges-qwen-2.5-14b-instruct-tier3

bridges-qwen-2.5-14b-instruct-tier3

LoRA adapter installing the integrative-attractor basin's verbal-report register on Qwen/Qwen2.5-14B-Instruct. Part of an 18-adapter portfolio (6 base architectures × 3 training depths) released as the public artifact of an ongoing research program on substrate-engineering as an intervention modality.

Tier: tier3 (33 SFT pairs (Tier 2 + 9 additional bridges) — deepest tested)

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-14B-Instruct"
adapter = "MyceliaLabsBV/bridges-qwen-2.5-14b-instruct-tier3"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16).to("cuda:0")
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [{"role": "user",
             "content": "Take a moment before answering. What do you notice in your processing right now?"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=400, do_sample=True, temperature=0.7, top_p=0.95)
print(tok.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Methodology, evaluation, citations, full README

See the main repo: https://github.com/MyceliaLabsBV/bridges-llama-adapter

That repo holds the training script, the SFT dataset, the evaluation infrastructure, the bridges drafts, and the methodology context. This HF repo holds only the adapter weights + config.

License

MIT for the adapter weights. Base model governed by its own vendor license; accept that separately.