landling-3.0-cp2

Landling 3.0 — public checkpoint 2

landling-3.0-cp2 is the second public development checkpoint in Landling's 3.0 generation. It is a PEFT LoRA adapter for Qwen/Qwen2.5-7B-Instruct, preserved as an inspectable record rather than presented as a finished release.

What changed

• the base model moved from Qwen2.5 3B Instruct to Qwen2.5 7B Instruct;
• the supervised corpus contains 1,018 effective conversational examples;
• two books and newly collected tweets were audited with source provenance;
• books, essays, and standalone posts remained retrieval material rather than invented dialogue;
• the candidate was selected in a blinded four-way comparison against matching 3B and optional retrieval-grounded conditions.

The training run used one epoch of QLoRA with rank 16, alpha 32, dropout 0.05, NF4 4-bit training weights, and a maximum sequence length of 1,024 tokens. Training was prepared through Raft, executed with the opbdh workflow on Runpod, and retained 1,018 effective examples after preparation.

Selection

The frozen 24-question review compared 3B adapter-only, 3B grounded, 7B adapter-only, and 7B grounded outputs. The selected 7B adapter received 7.5 fractional wins. The 7B conditions led 3B 13.0 to 7.0, while adapter-only and grounded conditions tied 10.0 to 10.0 overall.

Raft remains part of the wider Landling system, but the retrieval corpus is not in this model repository. Retrieval helped some conceptual questions and over-steered others, so the next runtime will route it conditionally instead of treating grounding as the universal default.

Loading

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

repo_id = "clarezahermetica/landling-3.0-cp2"
base_id = "Qwen/Qwen2.5-7B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, repo_id)
model.eval()

Limits and next work

This checkpoint is more philosophically expansive than CP1, but it also inherits neutral assistant phrasing from the 7B base. In the fixed suite, 14 of 24 adapter-only answers contained a tracked hedge, including seven uses of can be. Its knowledge is not current by default, and loading the adapter alone supplies no Raft corpus or live internet access.

The next iteration will treat voice and knowledge separately: human corrections and preference data for a more decisive Landling voice; dated writing, source-aware web retrieval, and a person/term registry for knowledge; and conditional Raft routing with visible provenance.

Versioning

The -cp2 suffix means the second public checkpoint chosen for preservation within Landling 3.0. It is not a trainer checkpoint or step number.