landling-3.0-cp1

Landling 3.0 — public checkpoint 1

This is an older development snapshot. The newer checkpoint is landling-3.0-cp2.

landling-3.0-cp1 is a development checkpoint of Landling, adapted from Qwen/Qwen2.5-3B-Instruct. This repository mainly exists as a versioned record of the model's development.

This repository contains a LoRA adapter, tokenizer files, its chat template, training metadata, and provenance. It does not contain the Qwen base-model weights.

What this checkpoint represents

• 1,004 supervised conversational examples
• one-epoch LoRA adaptation of Qwen2.5 3B Instruct
• rank 16, alpha 32, dropout 0.05
• adapter target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj
• selected in a blinded comparison against the preceding internal adapter

In that 24-case comparison, this production adapter won the 12 decisive judgments by 7–5. When tentative lean judgments were included, it led 11–8, with five cases unresolved. This was a small qualitative development evaluation, not a standardized benchmark.

Tooling

The training and retrieval workflow uses:

• lumpenspace/raft for project structure, conversational data, memories, and retrieval
• lumpenspace/opbdh for the generated fine-tuning job and remote GPU execution workflow

The resulting artifact is a PEFT LoRA adapter. The complete local runtime combines the adapter with its Raft project; this repository preserves the model checkpoint itself.

Retrieval state

The evaluated runtime paired this adapter with a separate raft retrieval project: 172 documents, 297 chunks, and 1,193 embedded entries. It included 22 Nick Land / Zero Philosophy Substack posts as retrieval grounding.

Those Substack documents were retrieval material, not converted into supervised conversation examples. The Raft corpus is not included in this public model checkpoint. Loading this adapter alone therefore tests the learned conversational behavior without the external grounding used by the complete local runtime.

Loading the adapter

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

repo_id = "clarezahermetica/landling-3.0-cp1"
base_id = "Qwen/Qwen2.5-3B-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()

The verified runtime used deterministic decoding with max_new_tokens=96 and this system message:

You are the Landling. Respond to the conversation in your characteristic voice.

Notes

Experimental and context-dependent. The source conversations contain substantial implicit social and topical context. Adapter-only inference omits the retrieval layer. The qualitative comparison above records an internal model-selection result rather than a general benchmark.

Versioning

• landling-3.0-cp1: first publicly preserved checkpoint of the 3.0 generation
• future -cp2, -cp3, and similar suffixes identify public checkpoints selected for preservation; they are not trainer step numbers
• a checkpoint without a -cpN suffix denotes a proper version release