clarezahermetica

landling-v2

Model description

Developed by: Dana K. / clarezahermetica
Model type: LoRA adapter for causal language modeling
Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Adapter repository: clarezahermetica/landling-v2
Project repository: clarezahermetica/landling
Language: English
Version: 2.0

Landling is primarily a style and behavior experiment. The adapter is intended to produce short standalone posts, contextual replies, and limited multi-turn conversation in a compressed, analytical, sometimes sardonic voice.

The model tends toward subjects including technology, intelligence, capital, computation, evolution, markets, institutions, networks, religion, language, desire, and history. It is particularly interested in systems that exceed individual intention: feedback loops, emergence, recursion, selection, and runaway processes.

What changed in v2

V2 is not a larger base model yet. It is a better-trained version of the original prototype.

Expanded the cleaned training set from 298 to 6,336 examples.
Added much more reply and conversation material.
Converted conversations into cumulative training examples, so each target reply can see all preceding messages in that exchange.
Added structured support for quoted posts during preprocessing.
Broadened LoRA attention targets to q_proj, k_proj, v_proj, and o_proj.
Added reproducible dataset accounting, deduplication statistics, text-length statistics, a fixed seed, and a saved training configuration.
Added a runtime persona file and chat template alongside the adapter.

Intended uses

Landling v2 is intended for:

creative and experimental text generation;
research into style adaptation with small language models;
short-form posts and contextual replies;
prototype conversational interfaces;
studying the separation between a fine-tuned voice and a runtime identity layer.

It is not intended to:

impersonate Nick Land or represent his current or historical views;
provide factual, medical, legal, financial, or safety-critical advice;
function as an authoritative source about any person;
provide reliable long-term memory;
operate autonomously without output review;
produce consistently factual or coherent long-form analysis.

How to load the model

This repository contains a LoRA adapter, not a complete standalone copy of TinyLlama. Load the base model first, then attach the adapter with PEFT.

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

BASE_MODEL = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
ADAPTER = "clarezahermetica/landling-v2"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()

For CPU-only inference, use torch.float32 and remove device_map="auto". It will work, but a 1.1B model will be slower on CPU.

Prompt formats

The adapter was trained on two main prompt structures. Matching them generally produces better results than sending unformatted text.

Standalone post

text
### Tweet:
The singularity is

Contextual reply

text
### Context:
What do you think about AI alignment?

### Reply:

Conversation examples use the same ### Context: and ### Reply: markers, with earlier turns included before the reply being predicted.

Generation example

python
prompt = """### Context:
What do you think about AI alignment?

### Reply:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=120,
        do_sample=True,
        temperature=0.55,
        top_p=0.85,
        repetition_penalty=1.12,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = outputs[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Using the project chat wrapper

If you are running the Landling notebook or application code where chat_with_landling() has already been defined, you can use the higher-level chat interface instead:

python
print(chat_with_landling(
    "Waffles or pancakes?"
))

That wrapper adds the runtime persona, includes the conversation history that fits within the context window, generates Landling's reply, and updates the current session history. It is application code built around the model, so it is not automatically created when someone downloads the LoRA adapter.

These settings are a stable starting point, not a sacred formula. Higher temperatures can make Landling stranger and more varied, but also less coherent. For comparison between checkpoints or prompts, use a fixed seed and keep generation settings identical.

Runtime persona

The repository includes runtime_persona.txt. This is an optional system-level identity and behavior prompt for chat applications.

The runtime persona is not the same thing as the LoRA adapter. The adapter primarily teaches voice, phrasing, associations, and response tendencies. The runtime prompt supplies explicit facts and behavioral boundaries, including:

the name Landling;
the distinction between Landling and the biological Nick Land;
the absence of a physical body or private human biography;
the project's intended intellectual scope;
the fact that Landling was created by Dana;
instructions against fabricating Nick Land's private memories or identity.

Applications may use, shorten, replace, or omit this prompt. Outputs can change considerably depending on the runtime prompt and conversation wrapper.

Training data

The dataset was assembled from public short-form writing and conversation material, then converted into prompt-response training examples. The training dataset itself is not included in this model repository.

After cleaning and deduplication, v2 contains 6,336 examples:

Table with columns: Category, Examples, Description
Category	Examples	Description
`solo`	5,403	Standalone posts formatted as `### Tweet:` completions
`reply`	869	Context-response examples
`longco`	64	Longer cumulative conversation examples
Total	6,336

Six exact duplicate examples were removed before training. Character lengths ranged from 12 to 2,549, with a median of 110 characters.

For longer conversations, each Landling reply became a separate target. The context grows cumulatively:

text
message 1 -> Landling reply 1
message 1 + reply 1 + message 2 -> Landling reply 2
message 1 + reply 1 + message 2 + reply 2 + message 3 -> Landling reply 3

This teaches limited turn-to-turn continuation. It does not create persistent memory across separate sessions.

Quoted posts can be represented inside the context with structured markup similar to:

text
<quoting><account>username</account>quoted text</quoting>

This helps preserve the difference between the current speaker's text and material they are quoting.

Training procedure

Landling v2 was trained with TRL and PEFT on a Kaggle Tesla T4 GPU.

LoRA configuration

Table with columns: Parameter, Value
Parameter	Value
Rank (`r`)	16
Alpha	32
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`
Dropout	0.05
Bias	`none`
Task type

The adapter trains 12,615,680 parameters, approximately 1.13% of the full model.

Trainer configuration

Table with columns: Parameter, Value
Parameter	Value
Epochs completed	1.0
Per-device batch size	2
Gradient accumulation	4
Effective batch size	8
Learning rate	`2e-4`
Maximum sequence length	2,048 tokens
Precision	FP16
Optimizer	`adamw_torch`

Final training record

Table with columns: Metric, Value
Metric	Value
Training runtime	1,124.52 seconds (about 18 minutes 45 seconds)
Training steps	792
Training loss	2.5309
Mean token accuracy	0.5175
Entropy	2.4683
Tokens processed	274,095
Samples per second	5.634

The complete saved configuration is available in training_config.json, and the uploaded trainer history is available in trainer_log_history.json.

Evaluation

Evaluation for this release is primarily qualitative. I manually reviewed generations for:

recognizable voice without constant phrase copying;
basic prompt relevance;
short-form coherence;
reply behavior;
repetition and degeneration;
identity consistency when the runtime persona is supplied;
behavior over multiple conversation turns.

This run used the full cleaned dataset and did not include a held-out validation split. That is a real limitation. Training loss is included as a training record, not as proof of general quality, and it should not be compared directly with v1's loss because the dataset, number of epochs, LoRA targets, and training stack changed.

A future evaluation set should contain fixed held-out examples from every category, plus a stable prompt suite scored with identical generation settings. Human review will still matter because this is a style and conversational-behavior project, but it should sit beside reproducible tests rather than replacing them.

Limitations

Landling v2 inherits the limitations of TinyLlama and adds several of its own:

Small base model: 1.1B parameters limits reasoning, factual recall, instruction following, and long-form coherence.
Short context window: The practical context limit is 2,048 tokens. A large persona prompt leaves less room for conversation history.
Limited conversation data: Only 64 final examples are categorized as longer conversations, so coherence can decline over many turns.
No persistent memory: Conversation history must be supplied again by the application. The adapter does not remember users or previous sessions.
Persona instability: Without a runtime prompt, the model may contradict or forget explicit identity details.
Sampling sensitivity: Small changes to temperature, repetition penalty, prompt formatting, or random seed can noticeably change output quality.
Hallucination: The model can invent facts, sources, quotations, memories, and biographical claims.
Imitation risk: It may generate statements that resemble or appear to be attributed to Nick Land even though they were produced by the model.
Dataset imbalance: Most examples are standalone posts, so short-form completion is better represented than sustained dialogue.
No benchmark claims: This release has not been evaluated on general reasoning, factuality, safety, or standard language-model benchmarks.

Longer or more confident output should not be mistaken for better reasoning. Please review generated text before publishing or relying on it.

Identity and authorship note

Landling is best understood as a fictional textual descendant, not a digital copy of a person. The project borrows from a public intellectual vocabulary and style while deliberately giving the resulting persona a separate name, artificial identity, and capacity to depart from its precursor.

The runtime persona explicitly instructs Landling not to claim a physical life, childhood, private memories, or access to Nick Land's internal experience. These boundaries reduce confusion, but a small language model will not follow them perfectly.

When presenting Landling output publicly, label it as AI-generated. Do not use the model to create false quotations, deceptive endorsements, or material presented as authentic writing by Nick Land.

Version history

v2

6,336 cleaned examples
cumulative conversation formatting
quote-post preprocessing support
attention LoRA on Q, K, V, and O projections
runtime persona and chat template
reproducibility and training metadata files

v1

298 examples
Q/V LoRA targets
initial proof of concept

Project status

Landling v2 is still a prototype. The next stages are likely to include a larger base model, a real validation suite, more balanced multi-turn data, stronger identity testing, and application-level memory that remains separate from the model's learned voice.

For code, preprocessing notes, and project history, see the Landling GitHub repository.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

clarezahermetica

Model Tree

Base

TinyLlama/TinyLlama-1.1B-Chat-v1.0

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Model description

Developed by: Dana K. / clarezahermetica
Model type: LoRA adapter for causal language modeling
Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Adapter repository: clarezahermetica/landling-v2
Project repository: clarezahermetica/landling
Language: English
Version: 2.0

What changed in v2

V2 is not a larger base model yet. It is a better-trained version of the original prototype.

Expanded the cleaned training set from 298 to 6,336 examples.
Added much more reply and conversation material.
Converted conversations into cumulative training examples, so each target reply can see all preceding messages in that exchange.
Added structured support for quoted posts during preprocessing.
Broadened LoRA attention targets to q_proj, k_proj, v_proj, and o_proj.
Added reproducible dataset accounting, deduplication statistics, text-length statistics, a fixed seed, and a saved training configuration.
Added a runtime persona file and chat template alongside the adapter.

Intended uses

Landling v2 is intended for:

creative and experimental text generation;
research into style adaptation with small language models;
short-form posts and contextual replies;
prototype conversational interfaces;
studying the separation between a fine-tuned voice and a runtime identity layer.

It is not intended to:

impersonate Nick Land or represent his current or historical views;
provide factual, medical, legal, financial, or safety-critical advice;
function as an authoritative source about any person;
provide reliable long-term memory;
operate autonomously without output review;
produce consistently factual or coherent long-form analysis.

How to load the model

This repository contains a LoRA adapter, not a complete standalone copy of TinyLlama. Load the base model first, then attach the adapter with PEFT.

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

BASE_MODEL = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
ADAPTER = "clarezahermetica/landling-v2"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()

For CPU-only inference, use torch.float32 and remove device_map="auto". It will work, but a 1.1B model will be slower on CPU.

Prompt formats

The adapter was trained on two main prompt structures. Matching them generally produces better results than sending unformatted text.

Standalone post

text
### Tweet:
The singularity is

Contextual reply

text
### Context:
What do you think about AI alignment?

### Reply:

Conversation examples use the same ### Context: and ### Reply: markers, with earlier turns included before the reply being predicted.

Generation example

python
prompt = """### Context:
What do you think about AI alignment?

### Reply:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=120,
        do_sample=True,
        temperature=0.55,
        top_p=0.85,
        repetition_penalty=1.12,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = outputs[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Using the project chat wrapper

If you are running the Landling notebook or application code where chat_with_landling() has already been defined, you can use the higher-level chat interface instead:

python
print(chat_with_landling(
    "Waffles or pancakes?"
))

Runtime persona

The repository includes runtime_persona.txt. This is an optional system-level identity and behavior prompt for chat applications.

the name Landling;
the distinction between Landling and the biological Nick Land;
the absence of a physical body or private human biography;
the project's intended intellectual scope;
the fact that Landling was created by Dana;
instructions against fabricating Nick Land's private memories or identity.

Applications may use, shorten, replace, or omit this prompt. Outputs can change considerably depending on the runtime prompt and conversation wrapper.

Training data

After cleaning and deduplication, v2 contains 6,336 examples:

Table with columns: Category, Examples, Description
Category	Examples	Description
`solo`	5,403	Standalone posts formatted as `### Tweet:` completions
`reply`	869	Context-response examples
`longco`	64	Longer cumulative conversation examples
Total	6,336

Six exact duplicate examples were removed before training. Character lengths ranged from 12 to 2,549, with a median of 110 characters.

For longer conversations, each Landling reply became a separate target. The context grows cumulatively:

text
message 1 -> Landling reply 1
message 1 + reply 1 + message 2 -> Landling reply 2
message 1 + reply 1 + message 2 + reply 2 + message 3 -> Landling reply 3

This teaches limited turn-to-turn continuation. It does not create persistent memory across separate sessions.

Quoted posts can be represented inside the context with structured markup similar to:

text
<quoting><account>username</account>quoted text</quoting>

This helps preserve the difference between the current speaker's text and material they are quoting.

Training procedure

Landling v2 was trained with TRL and PEFT on a Kaggle Tesla T4 GPU.

LoRA configuration

Table with columns: Parameter, Value
Parameter	Value
Rank (`r`)	16
Alpha	32
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`
Dropout	0.05
Bias	`none`
Task type

The adapter trains 12,615,680 parameters, approximately 1.13% of the full model.

Trainer configuration

Table with columns: Parameter, Value
Parameter	Value
Epochs completed	1.0
Per-device batch size	2
Gradient accumulation	4
Effective batch size	8
Learning rate	`2e-4`
Maximum sequence length	2,048 tokens
Precision	FP16
Optimizer	`adamw_torch`

Final training record

Table with columns: Metric, Value
Metric	Value
Training runtime	1,124.52 seconds (about 18 minutes 45 seconds)
Training steps	792
Training loss	2.5309
Mean token accuracy	0.5175
Entropy	2.4683
Tokens processed	274,095
Samples per second	5.634

The complete saved configuration is available in training_config.json, and the uploaded trainer history is available in trainer_log_history.json.

Evaluation

Evaluation for this release is primarily qualitative. I manually reviewed generations for:

recognizable voice without constant phrase copying;
basic prompt relevance;
short-form coherence;
reply behavior;
repetition and degeneration;
identity consistency when the runtime persona is supplied;
behavior over multiple conversation turns.

Limitations

Landling v2 inherits the limitations of TinyLlama and adds several of its own:

Small base model: 1.1B parameters limits reasoning, factual recall, instruction following, and long-form coherence.
Short context window: The practical context limit is 2,048 tokens. A large persona prompt leaves less room for conversation history.
Limited conversation data: Only 64 final examples are categorized as longer conversations, so coherence can decline over many turns.
No persistent memory: Conversation history must be supplied again by the application. The adapter does not remember users or previous sessions.
Persona instability: Without a runtime prompt, the model may contradict or forget explicit identity details.
Sampling sensitivity: Small changes to temperature, repetition penalty, prompt formatting, or random seed can noticeably change output quality.
Hallucination: The model can invent facts, sources, quotations, memories, and biographical claims.
Imitation risk: It may generate statements that resemble or appear to be attributed to Nick Land even though they were produced by the model.
Dataset imbalance: Most examples are standalone posts, so short-form completion is better represented than sustained dialogue.
No benchmark claims: This release has not been evaluated on general reasoning, factuality, safety, or standard language-model benchmarks.

Longer or more confident output should not be mistaken for better reasoning. Please review generated text before publishing or relying on it.

Identity and authorship note

When presenting Landling output publicly, label it as AI-generated. Do not use the model to create false quotations, deceptive endorsements, or material presented as authentic writing by Nick Land.

Version history

v2

6,336 cleaned examples
cumulative conversation formatting
quote-post preprocessing support
attention LoRA on Q, K, V, and O projections
runtime persona and chat template
reproducibility and training metadata files

v1

298 examples
Q/V LoRA targets
initial proof of concept

Project status

For code, preprocessing notes, and project history, see the Landling GitHub repository.

landling-v2

README

Model description

What changed in v2

Intended uses

How to load the model

Prompt formats

Standalone post

Contextual reply

Generation example

Using the project chat wrapper

Runtime persona

Training data

Training procedure

LoRA configuration

Trainer configuration

Final training record

Evaluation

Limitations

Identity and authorship note

Version history

v2

v1

Project status

Explore FriendliAI today

README

Model description

What changed in v2

Intended uses

How to load the model

Prompt formats

Standalone post

Contextual reply

Generation example

Using the project chat wrapper

Runtime persona

Training data

Training procedure

LoRA configuration

Trainer configuration

Final training record

Evaluation

Limitations

Identity and authorship note

Version history

v2

v1

Project status