Rnoooo/cvd-lora

astertech-cvd-coach-lora-r0

LoRA adapter for Qwen/Qwen2.5-VL-7B-Instruct that performs streaming step-completion detection on egocentric video of CVD (chemical vapour deposition) tube-furnace experiments. At each tick the model is shown the last 30 s of footage plus the current instruction and answers a Yes/No question about whether that step is now complete, so a coach can acknowledge the step and advance.

This is round 0 of the AsterTech-vLLM project (milestone M3). Completion detection only — mistake detection is deliberately out of scope.

Provenance and access

Trained on local-only laboratory footage that is not distributed and never leaves the project's own machines. The weights here are derived from that footage, which is why this repository is private. Treat any change of visibility as a deliberate decision about derived data, not a routine toggle.

Training recipe

Reproduced verbatim from results/m3/pooled_provenance.json (seed 0, deterministic inputs: hash-pinned annotations + frozen benchmark manifest benchmarks/v1.json).

Base model	Qwen/Qwen2.5-VL-7B-Instruct (bf16)
LoRA	r=16, α=32, dropout=0.05
Optimiser	lr 1e-5, batch 1 × grad-accum 8, 1 epoch, seed 0
Frames	2 fps, ≤16 frames/example, 30 s clip window, max_pixels 151200
Pair generation	completion_pairs_from_gt(short_yes=True), neg_per_step=4, neg_margin_s=8, min_step_s=12
Dataset	672 examples — 448 Yes / 224 No (56 neg-eligible steps ×5 + 392 yes-only + 3 zero-length skips)
Train videos	the 13 pool videos of benchmark v1 (the 3 frozen holdout videos are never trained on)

Training data is same-step contrastive pairs: for each annotated step one "Yes" window ending at the completion time and N "No" windows ending mid-step, with identical instruction text and window length, so only the pixels differ. Every example comes from one code path over the project's own ground truth — no external corpora, no mixed sources.

Results (M3, benchmark v1)

Evaluated at the frozen v1 eval config (tick 5 s / window 30 s / 2 fps / ≤16 frames / confirm 1; match window 30 s full width).

split	zero-shot IC-Acc	SFT IC-Acc	delta	missed	overtalk
pool (LOVO, 13 folds)	14.63% (66/451)	48.78% (220/451)	+34.15	0.854 → 0.512	0.283 → 0.315
frozen holdout (this pooled adapter)	20.48% (17/83)	54.22% (45/83)	+33.74	0.795 → 0.458	0.292 → 0.262

The holdout videos were never trained on in any run, and gained as much as the LOVO folds — which is what rules out "it memorised the training videos".

Slice movement on the pool (zero-shot → SFT): short steps 8/80 → 49/80, pressure waits 1/64 → 18/64, other 57/311 → 154/311.

About this copy — a reproduction, not the original artifact

The original M3 pooled adapter was written to instance-store disk and lost when the training box was stopped. These weights are a re-run of the identical recipe on 2026-08-04: same committed annotations, same hash-pinned manifest, same 13 train ids, seed 0, same hyperparameters. The reproduction is exact on every input we can check — the dataset came out at 672 examples / 448 Yes / 224 No, matching the original pin byte-for-byte, and training took 66.9 min against the original's 67.0 min, ending at train_loss 0.4088.

GPU nondeterminism means the weights are not bit-identical to the originals, so the table above is the original run's measurement. See "Verification" below for this copy's own measured holdout score.

Caveats — read these before trusting it

1. Overtalk rose on the pool (0.283 → 0.315). The 2:1 Yes bias buys detections partly through more emissions. Holdout overtalk improved (0.292 → 0.262), so this is not a degenerate always-yes model, but it is a trade. The next lever is harder mid-step negative windows.
2. The pressure-wait gain is not gauge reading. The DigiVac LCD is provably illegible at the model's 360×420 effective input, so the model is learning the operator's reaction cue (stillness, then reaching for the next valve), not the displayed value. Do not present it as instrument reading.
3. Small dataset. 16 videos / 534 completions total. Single-video scores (n = 21–49) are noise; only the aggregates above carry signal.

Verification

These exact weights were re-scored on the 3 frozen holdout videos (2026-08-04), loaded from this repository by id — so the number below also verifies the download path, not just the local files.

metric	original M3 adapter	this copy
IC-Acc	54.22% (45/83)	57.83% (48/83)
missed	0.458	0.422
overtalk	0.262	0.238
mean signed offset	—	+1.02 s (abs 4.19 s)

+3.61 pts against a ±5 pt reproduction gate — pass. The difference is GPU nondeterminism at n=83 (3 completions), not a better model; both are the same recipe and the original table is the one to quote. Zero-shot on the same 3 videos is 20.48% (17/83).

Usage

from peft import PeftModel
from transformers import AutoModelForVision2Seq

base = AutoModelForVision2Seq.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", dtype="bfloat16")
model = PeftModel.from_pretrained(base, "Rnoooo/astertech-cvd-coach-lora-r0")

Requires authentication (HF_TOKEN) while the repository is private. The adapter expects the exact Yes/No completion prompt shape it was trained on — the project's astertech.prompts.completion_messages. Using a different prompt string collapses recall; that was the central bug of an earlier round.