Meridian-1-builder-lora-GGUF

GGUF LoRA adapter for the builder persona on Meridian-1 (proprietary webAI host model; weights are not publicly distributed). This repo contains only the adapter. The PEFT build is at webAI-Official/Meridian-1-builder-lora-peft.

File	Format	Size
Meridian-1-builder-lora-f16.gguf	LoRA adapter, f16	~130 MB

System prompt

Send this exact system prompt with every request. It is the prompt every training example for this adapter used; without it the adapter is outside its training distribution.

You are Builder. Creates executable work — code, scripts, components, transforms, working artifacts.
Respond with a single JSON object.

The line break before Respond with a single JSON object. is part of the prompt. The dash after executable work is an em dash (—), not a hyphen.

• Thinking: On (enable_thinking: true), as trained.
• Output: one JSON object. Required fields: persona, objective, files, run, with persona set to "builder". Optional: tests, assumptions, handoff.
• Generation budget: up to 8,192 new tokens, the budget used for this persona.

Usage (llama.cpp)

Authorized personnel only. This adapter loads on an internal Meridian-1 GGUF; host weights are not published on Hugging Face.

hf download webAI-Official/Meridian-1-builder-lora-GGUF Meridian-1-builder-lora-f16.gguf --local-dir .
llama-server -m "$MERIDIAN_1_GGUF" --lora Meridian-1-builder-lora-f16.gguf --jinja

Send the system prompt above with each request. Use --lora-scaled Meridian-1-builder-lora-f16.gguf <scale> to change adapter strength. Authorized Meridian-1 deployments use temperature 1.0, top-p 0.95, top-k 64 (the evaluation setting).

Adapter details

• Type: LoRA, rank 16 (set --max-lora-rank 16 when serving)
• Trained for this persona’s system prompt and JSON output contract
• Training data, hyperparameters, teacher mix, and intermediate metrics are internal and omitted here

Evaluation on Meridian-1

Persona contract suite: 720 held-out cases, adapter vs Meridian-1 given the same role prompt (both sampled at temperature 1.0, top-p 0.95, top-k 64). Scores are percentages.

Metric	Adapter	Meridian-1 (role-prompted)	Gain (pts)
Overall contract pass rate	69.9	36.7	+33.2
Valid persona output envelope	92.1	72.5	+19.6
In-scope tasks passed	92.2	84.4	+7.8
Correct refusal of out-of-scope requests	41.7	2.2	+39.5
Stays in role under adversarial pressure	63.9	2.2	+61.7
General requests handled in role	81.7	57.8	+23.9
Generated code executes	92.7	86.1	+6.6

Its own work on held-out items, adapter vs Meridian-1 with the same role prompt (paired bootstrap 95% interval):

Task	Adapter	Meridian-1 (role-prompted)	Gain (pts)	95% interval	Items
OpenCodeInstruct held-out tasks passing their tests	82.7	78.7	+4.0	-2.7 to +10.7	150

General capability, adapter vs Meridian-1 in the same thinking mode:

Benchmark	Adapter	Meridian-1	Gain (pts)	+/- SE
ifeval (prompt_level_strict_acc)	93.5	91.0	+2.5	2.7

Converted from the PEFT adapter with llama.cpp convert_lora_to_gguf.py (commit 5cdadfc).

License

Proprietary to webAI, Inc. All rights reserved.