| # Meta Llama 3.1 8B Instruct — Local Baseline Worker |
|
|
| **Status:** optional software contract implemented; no new Llama 3.1 inference has been executed by FrontierLab. |
|
|
| ## Why it joins |
|
|
| `meta-llama/Llama-3.1-8B-Instruct` fills the middle of FrontierLab's model-size spectrum. |
|
|
| Meta released Llama 3.1 on July 23, 2024. The 8B Instruct model is text-only, supports a 128K-token context, eight officially supported languages, Grouped-Query Attention, multilingual dialogue and tool-use-oriented workflows. |
|
|
| Official references: |
|
|
| - https://ai.meta.com/blog/meta-llama-3-1/ |
| - https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct |
|
|
| FrontierLab does **not** add it because it is newer or stronger than Muse Glimmer. It joins as a mature, widely deployed medium-size baseline. |
|
|
| ## Role |
|
|
| The intended role is: |
|
|
| **Local General-Purpose Baseline / Worker** |
|
|
| It can run the same structured `json(role, task, payload)` interface used by FrontierLab research agents, but every receipt is permanently marked: |
|
|
| ```text |
| baseline_only = true |
| scientific_output_independently_verified = false |
| ``` |
|
|
| This makes it suitable for controlled A/B comparisons while keeping the verified Glimmer path distinct. |
|
|
| ## Future scale-vs-reliability comparison |
|
|
| A useful controlled comparison is: |
|
|
| ```text |
| Muse2-230M -> edge structured preprocessing |
| Llama 3.1 8B -> local general-purpose baseline |
| Muse Glimmer 30B -> verified scientific reasoning backend |
| Muse Spark 816B -> frontier multimodal reasoning |
| ``` |
|
|
| Run the same bounded structured task and compare: |
|
|
| - JSON validity; |
| - abstention rate; |
| - citation-ID validity; |
| - exact quotation validity; |
| - researcher intervention rate; |
| - latency; |
| - input/output size or token usage. |
|
|
| The goal is not to prove that larger models are always better. It is to identify the smallest model that can reliably perform each bounded job. |
|
|
| ## Runtime boundary |
|
|
| The default FrontierLab installation remains unchanged. No 8B checkpoint, Transformers runtime or GPU dependency is added to `requirements.txt`. |
|
|
| `Llama31BaselineClient` accepts an injected local runtime exposing: |
|
|
| ```python |
| runtime.complete( |
| messages=[...], |
| max_new_tokens=1024, |
| temperature=0.2, |
| top_p=0.9, |
| ) |
| ``` |
|
|
| This allows llama.cpp, Transformers or another local serving layer to be wrapped without coupling FrontierLab to one runtime. |
|
|
| ## Scientific boundary |
|
|
| Llama 3.1 8B can participate in a controlled baseline version of evidence extraction, skeptical review or hypothesis formatting, but its output still goes through the same: |
|
|
| ```text |
| source ID -> exact quote -> human review -> evaluation -> planning gate |
| ``` |
|
|
| It never receives authority to approve its own scientific output. |
|
|
| Run the zero-network preflight: |
|
|
| ```bash |
| python -m scripts.llama31_baseline_preflight |
| ``` |
|
|
| No checkpoint is downloaded by that command. |
|
|
