TinyLlama-1.1B-HolyC-Layer2

TinyLlama-1.1B-HolyC Layer 2

Layer 2 is the completion-oriented adapter in the stack. If layer 1 is the reader and explainer, layer 2 is the one that rolls up its sleeves and starts writing HolyC back at you.

What This Layer Adds

Compared with layer 1, this adapter is better positioned for:

• function completion from a signature or partial prefix
• fill-in-the-middle reconstruction of HolyC code
• comment-to-code and spec-to-code style prompting
• code generation tasks that benefit from seeing HolyC as executable structure rather than only text to explain

Provenance

This layer needs one careful note.

• The cycle 2 training code loads Aptlantis/TinyLlama-1.1B-HolyC as the starting model.
• That means the run was operationally built on top of the published layer 1 adapter.
• The saved PEFT adapter metadata still names TinyLlama/TinyLlama-1.1B-Chat-v1.0 as base_model_name_or_path.

So consumers should treat the layer relationship as documented provenance. The stack is real, but it is not fully self-describing inside the exported adapter config.

Training Data

Layer 2 is tied to the completion dataset in this bundle:

• Aptlantis/holyC-tinyllama-two-layer: umbrella dataset bundle
• completion dataset: datasets/completions/holyC_completion.jsonl inside that bundle, 8683 records
• codebase dataset: datasets/codebase/holyC_codebase.jsonl inside that bundle, 3448 raw HolyC source records

The completion dataset includes several generation-oriented sample shapes:

• signature to full body
• partial prefix to continuation
• fill-in-the-middle reconstruction
• leading comments to implementation

These samples were generated from TempleOS HolyC source using the dataset-generation scripts documented in the Aptlantis/holyC-tinyllama-two-layer bundle.

How To Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
adapter_path = "./layer2"

tokenizer = AutoTokenizer.from_pretrained(adapter_path)
base_model = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base_model, adapter_path)

If you want to honor the documented lineage exactly, treat this adapter as the second stage in a chain rooted in layer 1. If you want the simplest PEFT loading path, use it directly on the TinyLlama base and keep the lineage note in mind.

Intended Use

Use layer 2 for:

• HolyC code completion experiments
• TempleOS-style function generation
• adapter stacking research and release archaeology
• playful but serious niche-code generation

Limitations

• This bundle ships the publishable adapter artifacts, not a merged full model.
• The adapter metadata does not fully express the layer-on-layer training history by itself.
• No strong claim is made here about benchmark superiority over layer 1; the main distinction is training objective and practical behavior.
• Generated code still needs human review, especially for low-level or system-adjacent routines.

Quick Characterization

Layer 1 explains the ritual. Layer 2 performs it.