eric-ml-nlp/Delveta-LayaChoice-v2-checkpoints

Delveta LayaChoice v2 — historical training checkpoints

Scope. This repository is intended to hold only the non-selected historical training checkpoints of the V2 run (epochs 1–6), plus a record of the full 10-epoch training history.

• The selected model is Epoch 7 and is published separately at eric-ml-nlp/Delveta-LayaChoice-v2.
• This repository is not the inference entry point. Use the selected model repository for inference.
• Epochs 1–6 are historical training artifacts. They are not independent evaluation results and are not claimed to be better than Epoch 7; Epoch 7 was selected because it had the highest validation top-1.

Publication status. The epoch 1–6 weight files are published in this repository (see Checkpoint Inventory and SHA256 checksums). Weight files for epochs 7–10 are intentionally not included — Epoch 7 is published as the selected model, and epochs 8–10 are recorded here as history only.

1. Repository Purpose

This repository exists so the training history of the V2 run is auditable: it records every epoch's training loss and validation metrics, identifies which checkpoint was selected, and (once staged) publishes the non-selected epoch weights that were kept.

It complements:

• the selected model (Epoch 7): https://huggingface.co/eric-ml-nlp/Delveta-LayaChoice-v2
• the frozen dataset: https://huggingface.co/datasets/eric-ml-nlp/Delveta-LayaChoice-v2-Data

2. Relationship to the Selected Model

Selected checkpoint	Epoch 7
Selection rule	best_epoch = argmax(validation top-1), ties resolved to the earlier epoch; train loss never used
Epoch-7 validation top-1	290/300 = 96.67 %
Selected model repository	https://huggingface.co/eric-ml-nlp/Delveta-LayaChoice-v2

Epochs 8–10 are not better than Epoch 7: their validation top-1 is 96.33 %, below Epoch 7's 96.67 %.

3. Checkpoint Inventory

The epoch 1–6 model.safetensors weights are published. Each is the raw fine-tuned weight file for that epoch; sizes and digests are the real published values. Epoch 7's validation top-1 (96.67 %) was the highest, so Epoch 7 was selected.

Epoch	Path	Size	Validation top-1
1	epoch-1/model.safetensors	1.29 GB	85.67%
2	epoch-2/model.safetensors	1.29 GB	90.00%
3	epoch-3/model.safetensors	1.29 GB	94.00%
4	epoch-4/model.safetensors	1.29 GB	89.67%
5	epoch-5/model.safetensors	1.29 GB	94.67%
6	epoch-6/model.safetensors	1.29 GB	94.33%

Epochs 7–10 weight files are intentionally out of scope for this repository: Epoch 7 is the selected model (published in the model repository) and Epochs 8–10 are recorded here as history only. No weight file from epochs 7–10 is placed here, and no epoch's file was copied or renamed to stand in for another epoch.

4. Validation Metrics by Epoch

Validation v3, 300 rows, card view B_noprov. Top-1 is invariant under the calibration temperature. This table records the training history; it does not imply this repository holds the weights of every epoch.

Epoch	Train Loss	Val Top-1	Val ECE	REJECT Recall	REJECT FPR
1	0.641706	85.67%	0.080275	81.25%	8.45%
2	0.384151	90.00%	0.036385	93.75%	8.10%
3	0.256366	94.00%	0.056254	93.75%	2.46%
4	0.132143	89.67%	0.101781	100.00%	8.10%
5	0.079365	94.67%	0.054404	100.00%	2.82%
6	0.039875	94.33%	0.056691	100.00%	2.82%
7	0.013119	96.67%	0.030319	100.00%	1.76%
8	0.000358	96.33%	0.035276	100.00%	2.11%
9	0.000000	96.33%	0.035276	100.00%	2.11%
10	0.000000	96.33%	0.035276	100.00%	2.11%

• Epoch 7 is the selected model (validation top-1 = 96.67 % = 290/300).
• Validation REJECT counts are small (16 gold REJECT rows per epoch), so per-epoch REJECT recall/FPR are noisy and should not be read as precise estimates.
• Epochs 9–10 are identical to Epoch 8 because the training loss had reached 0 and the weights stopped moving.

5. Training Provenance

Base model	convaiinnovations/laya
Base revision	55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851
Base subfolder	multilingual
Encoder	jhu-clsp/mmBERT-base
Method	full-parameter fine-tuning (no adapters)
Runtime	laya==0.3.21
Dataset	https://huggingface.co/datasets/eric-ml-nlp/Delveta-LayaChoice-v2-Data
Candidate count	K = 3
Card representation	B_noprov
Decision setup	four-way decision including REJECT

Optimizer details, exact hardware and wall-clock training time are not asserted here — they are not recorded in the artifacts this card was built from. See the selected model card for the training configuration that is recorded.

6. File Contents and Resumability

Each epoch-N/ directory is expected to hold a model.safetensors weight file. A weight file alone is not a fully resumable training checkpoint: full resumption requires the optimizer state, LR scheduler state, step counter and RNG state from the same step. Unless those are also published alongside the weights, this repository supports inference-style weight loading, not resumption of training from the mid-run state.

7. SHA256 Checksums

30f28e23cf8b3e0fa76fd7b1f74337b49d69b6eec3932e21ebc221314be8203f  epoch-1/model.safetensors
ea0f866aed715141733b25b00a643f46392b3a61230ee55e399ce1688866c2d2  epoch-2/model.safetensors
27dd2756ec2a7547555fcae2b2acf6d37cc10d456a132fc6754357f1487fcb5b  epoch-3/model.safetensors
e04b5375cf46273744d7a6e47698a258278735d74c0061eed227921fc80bc56f  epoch-4/model.safetensors
4572dd2d3986b58476685bdc9151458a308d15808904a54d9b09b39120d832dc  epoch-5/model.safetensors
a2f1046c5f8068ac13afe86516bc5ece7961eea410adabf590c1ca829218a207  epoch-6/model.safetensors

These are the SHA256 digests of the published weight files, cross-checked against the published repository's LFS metadata. Epoch 7's published weight is the same bytes as the model repository's model.safetensors (c6331bd5…).

8. Reproduction Notes

• To reproduce the selected model's behaviour, use Epoch 7 from Delveta-LayaChoice-v2, not a checkpoint from this repository.
• The training/validation metrics above were read from the run's per-epoch metric records, not reconstructed from memory.
• The dataset used for training is frozen and published (see §5).

9. Limitations

• Historical checkpoints are superseded by the selected Epoch 7 model and are provided for audit / research only.
• Validation REJECT metrics are computed over only 16 gold REJECT rows per epoch and are correspondingly uncertain.
• Validation top-1 is a 300-row estimate; small differences between epochs (e.g. 96.33 % vs 96.67 %) are within the noise of that split.
• No claim is made that any unselected epoch generalizes better than Epoch 7.

10. License

No license is asserted for these weights. The laya software package is Apache-2.0, but the license of the base model weights (convaiinnovations/laya) could not be determined at release time. Downstream use should confirm terms with the upstream rights holder.