v1 vLLM run specifications

Models, context, prompts, sampling, and decode settings for prompt-v1 baselines

Companion: llm_v2_v3_enhanced_variants.md · persona_prompt_versions.qmd · config/vllm_presets.yaml

This page pins the prompt-v1 digital-twin generation stack used for the four published full-cohort baselines (N = 241 × 5 = 1,205 prompts each). Primary manuscript claims remain on these runs.

1 Models (prompt v1)

Model	HF id	Role in v1 ranking
Llama-3.1-8B-Instruct	`meta-llama/Llama-3.1-8B-Instruct`	Cautionary transit IP collapse
Llama-3.2-3B-Instruct	`meta-llama/Llama-3.2-3B-Instruct`	Best group bands / IP MAE among non-collapsed
DeepSeek-R1-Distill-Llama-8B	`deepseek-ai/DeepSeek-R1-Distill-Llama-8B`	Best pooled group MAE; tier-stable
Llama-3.3-70B-Instruct	`meta-llama/Llama-3.3-70B-Instruct`	Mode-collapse cautionary case

Export stamps (where documented): Llama-3.1 20260726_0221; Llama-3.2 …_20260726_0252; DeepSeek …_20260726_0324. Host for 8B/3B/DeepSeek: rogers-gpu-1.discovery.wisc.edu (RTX A5000; fp8 Marlin weight-only where stated).

2 Prompt styling (v1)

Layer	v1 specification
Framing	AI Terrarium / ICA 2026 digital twin: second-person narrative (“You are a …”), not checklist / “adopt this profile”
System prompt	Inhabit persona; first-person CA self-report; JSON-only contract with group + interpersonal scores (6–30) and bands (low/moderate/high). Historical v1 system text omitted the later independent-subscale / non-deterministic / mobility-anti-bleed clauses
User prompt	Cumulative narrative (`demos` → `employment` → `geo` → `transit` → `full`) + fused CA ask
Geo	High-precision lat/lon (≈4 decimal)
Transit	Full mobility dump including rides-per-day even when frequency was Never
Instruction style	Instruct-tuned chat template (`tokenizer.apply_chat_template`); system + user messages; DeepSeek may emit `<think>` traces before JSON (post-processed at ingest)
JSON constraint	Soft prompt only (“Respond with ONLY a JSON object”); no guided decoding in v1

Current tree code implements v2/v3.1-enhanced packaging (signal-first + anti-bleed). Historical v1 wording is preserved in this page and the published baseline write-ups, not as a live --prompt-version switch.

3 Runtime / decode (v1)

Values below are the predict_vllm.py / run_vllm.sh defaults that applied when published launchers omitted sampling flags. Write-ups documented quantisation and throughput for some models but often omitted temperature; treat greedy defaults as the reproducible reconstruction.

Parameter	v1 value	Notes
`temperature`	0.0	Greedy decode (code default)
`top_p`	1.0	Nucleus disabled
`repetition_penalty`	1.0	Off
`seed`	unset	Not reproducible across engines
`max_output_tokens`	256	CA JSON headroom
`max_model_len` (context)	8192	Shell / CLI default
`quantization`	fp8	Documented for 3.1 / DeepSeek; assumed for siblings
`gpu_memory_utilization`	0.9
`batch_size` / `save_freq`	16 / 200	Checkpoint-resume CSV
`tensor_parallel_size`	1 (shell) / 2 (CLI default)	Published 8B runs used TP=1 on A5000
Guided JSON	false	Prompt-only contract

Preset name for reconstruction: v1_baseline in config/vllm_presets.yaml.

VLLM_PRESET=v1_baseline ./scripts/run_vllm.sh
# or
python -m inference.predict_vllm --preset v1_baseline ...

4 Why these settings matter for RQs

Greedy + soft JSON recovered valid JSON at 100% for Llama instruct models, but Llama-3.3-70B collapsed to ≈93% constant prior (18, 12) — parse success ≠ twin success.
Fused transit dump + high-precision geo coincided with Llama-3.1 interpersonal MAE 4.67 → 8.17 and signed error −2.2 → +6.5 (systematic high-IP over-prediction).
No seed limits exact replay of stochastic siblings; v2/v3 enhanced presets fix seed=42.

Enhanced decode + packaging for future re-runs: llm_v2_v3_enhanced_variants.md.

--- title: "v1 vLLM run specifications" subtitle: "Models, context, prompts, sampling, and decode settings for prompt-v1 baselines" --- **Companion:** [`llm_v2_v3_enhanced_variants.md`](llm_v2_v3_enhanced_variants.md) · [`persona_prompt_versions.qmd`](persona_prompt_versions.qmd) · [`config/vllm_presets.yaml`](../config/vllm_presets.yaml) This page pins the **prompt-v1** digital-twin generation stack used for the four published full-cohort baselines (N = 241 × 5 = 1,205 prompts each). Primary manuscript claims remain on these runs. --- ## Models (prompt v1) | Model | HF id | Role in v1 ranking | |---|---|---| | Llama-3.1-8B-Instruct | `meta-llama/Llama-3.1-8B-Instruct` | Cautionary transit IP collapse | | Llama-3.2-3B-Instruct | `meta-llama/Llama-3.2-3B-Instruct` | Best group bands / IP MAE among non-collapsed | | DeepSeek-R1-Distill-Llama-8B | `deepseek-ai/DeepSeek-R1-Distill-Llama-8B` | Best pooled group MAE; tier-stable | | Llama-3.3-70B-Instruct | `meta-llama/Llama-3.3-70B-Instruct` | Mode-collapse cautionary case | Export stamps (where documented): Llama-3.1 `20260726_0221`; Llama-3.2 `…_20260726_0252`; DeepSeek `…_20260726_0324`. Host for 8B/3B/DeepSeek: `rogers-gpu-1.discovery.wisc.edu` (RTX A5000; fp8 Marlin weight-only where stated). --- ## Prompt styling (v1) | Layer | v1 specification | |---|---| | **Framing** | AI Terrarium / ICA 2026 digital twin: second-person narrative (“You are a …”), not checklist / “adopt this profile” | | **System prompt** | Inhabit persona; first-person CA self-report; JSON-only contract with group + interpersonal scores (6–30) and bands (low/moderate/high). Historical v1 system text **omitted** the later independent-subscale / non-deterministic / mobility-anti-bleed clauses | | **User prompt** | Cumulative narrative (`demos` → `employment` → `geo` → `transit` → `full`) + fused CA ask | | **Geo** | High-precision lat/lon (≈4 decimal) | | **Transit** | Full mobility dump including rides-per-day even when frequency was Never | | **Instruction style** | Instruct-tuned chat template (`tokenizer.apply_chat_template`); system + user messages; DeepSeek may emit `<think>` traces before JSON (post-processed at ingest) | | **JSON constraint** | Soft prompt only (“Respond with ONLY a JSON object”); **no** guided decoding in v1 | Current tree code implements **v2/v3.1-enhanced packaging** (signal-first + anti-bleed). Historical v1 wording is preserved in this page and the published baseline write-ups, not as a live `--prompt-version` switch. --- ## Runtime / decode (v1) Values below are the **`predict_vllm.py` / `run_vllm.sh` defaults** that applied when published launchers omitted sampling flags. Write-ups documented quantisation and throughput for some models but often omitted temperature; treat greedy defaults as the reproducible reconstruction. | Parameter | v1 value | Notes | |---|---:|---| | `temperature` | **0.0** | Greedy decode (code default) | | `top_p` | **1.0** | Nucleus disabled | | `repetition_penalty` | **1.0** | Off | | `seed` | **unset** | Not reproducible across engines | | `max_output_tokens` | **256** | CA JSON headroom | | `max_model_len` (context) | **8192** | Shell / CLI default | | `quantization` | **fp8** | Documented for 3.1 / DeepSeek; assumed for siblings | | `gpu_memory_utilization` | **0.9** | | | `batch_size` / `save_freq` | **16** / **200** | Checkpoint-resume CSV | | `tensor_parallel_size` | **1** (shell) / **2** (CLI default) | Published 8B runs used TP=1 on A5000 | | Guided JSON | **false** | Prompt-only contract | Preset name for reconstruction: `v1_baseline` in [`config/vllm_presets.yaml`](../config/vllm_presets.yaml). ```bash VLLM_PRESET=v1_baseline ./scripts/run_vllm.sh # or python -m inference.predict_vllm --preset v1_baseline ... ``` --- ## Why these settings matter for RQs - **Greedy + soft JSON** recovered valid JSON at 100% for Llama instruct models, but **Llama-3.3-70B** collapsed to ≈93% constant prior `(18, 12)` — parse success ≠ twin success. - **Fused transit dump + high-precision geo** coincided with Llama-3.1 interpersonal MAE 4.67 → **8.17** and signed error −2.2 → **+6.5** (systematic high-IP over-prediction). - **No seed** limits exact replay of stochastic siblings; v2/v3 enhanced presets fix seed=42. Enhanced decode + packaging for future re-runs: [`llm_v2_v3_enhanced_variants.md`](llm_v2_v3_enhanced_variants.md).