foam-cfd-unified-14b-private (v3 cycle-2 candidate)

A private v3 candidate of the foam-cfd-unified-14b agent, produced by one cycle of the seven-layer self-evolution pipeline introduced in v3.

This checkpoint = base merged 14B (v2) $+$ one corrective QLoRA epoch (loss $0.35\to0.04$) on a 289-row training mix:

• 184 rows curated from the rolling capture file (score $\geq 0.65$)
• 79 rows from the frozen v1 anchor corpus (Layer-5 anchor mix at 30%)
• 26 rows of fresh active-learning prompts on the weakest solver family from the previous OOD eval (Layer-6 active learning, target family = pimpleFoam)

Why this is private

This is an internal candidate, not the production v2 model. Use the public foam-cfd-unified-14b unless you specifically need to A/B against the v3 cycle-2 changes documented below.

Behaviour vs the v2 baseline

On the held-out 110-prompt OOD evaluation in raw-LLM mode (agent-side keyword guards disabled):

Metric	v2 (public)	v3 cycle-2 (this)
End-to-end PASS rate	$110/110 = 100.0%$	$110/110 = 100.0%$
Solver-pick exact match	$106/110 = 96.4%$	$106/110 = 96.4%$
Same 4 borderline mismatches	yes	yes
Per-prompt FILES_CHANGE	n/a (baseline)	4 prompts (3 attributable to active-learning training, 1 noise floor)
Per-prompt regression flags	n/a	0

The aggregate eval gate would have promoted both models equally; the per-prompt regression diff (Layer 7) reveals 3 prompts whose case files differ as a result of the active-learning training, with no SUCCESS_FLIP, SOLVER_CHANGE, or SCORE_DROP.

How it was trained

• Base model: arungovindneelan/foam-cfd-unified-14b (v2 merged, bf16)
• Method: QLoRA via Unsloth, then bf16 merge for vLLM
• LoRA configuration: $r = 64$, $\alpha = 128$, dropout $= 0.0$, target $= {q,k,v,o,gate,up,down}_\text{proj}$
• Schedule: 1 corrective epoch on top of v2, paged AdamW 8-bit, bf16, batch $1 \times$ grad-accum $8$, max sequence length 8192, learning rate $2 \times 10^{-4}$, cosine schedule, 5% warm-up. Reward-weighted loss with weights $w_i = s_i^2 / \overline{s^2}$.
• Training data: 289-row mix described above
• Hardware: single H100 80GB
• Wall time: 5 min training $+$ 1 min adapter merge

Quick start

from vllm import LLM, SamplingParams
llm = LLM(
    model="arungovindneelan/foam-cfd-unified-14b-private",
    dtype="bfloat16",
    max_model_len=8192,
    gpu_memory_utilization=0.55,
)
out = llm.chat(
    [{"role": "system", "content": "You are an expert OpenFOAM CFD engineer..."},
     {"role": "user",   "content": "2D lid-driven cavity Re=1000, 2m square, water"}],
    sampling_params=SamplingParams(temperature=0.0, max_tokens=1024),
)
print(out[0].outputs[0].text)

Related artefacts

• Public v2 model: arungovindneelan/foam-cfd-unified-14b
• Public v2 deploy code: github.com/AGN000/foam-cfd-deploy
• Private v3 agent code: github.com/AGN000/foam-cfd-agent-v3-private
• Training corpus (anchor v1, 402 rows): arungovindneelan/openfoam-Agent-Dataset