Update model card for promoted 1-epoch weights

README.md

@@ -14,7 +14,27 @@ pipeline_tag: text-generation
 
 # Agron (domain-pretrained: programming languages + cybersecurity)
 
-Agron continued pretraining started from [`teddyoas/Agron`](https://huggingface.co/teddyoas/Agron) (published checkpoint @ commit [`a311e94`](https://huggingface.co/teddyoas/Agron/commit/a311e94867629620de8d0a45136acb5d0ba91755), ~50.3M params LLaMA-style).
+Agron is a ~50.3M-parameter LLaMA-style model. The weights in this repository are
+the result of **one full epoch** of domain continued pretraining over code,
+cybersecurity, and WikiText replay, started from the previously published Agron
+checkpoint.
+
+## Current promoted weights
+
+| Field | Value |
+|---|---|
+| Steps | 7,246 (1 full epoch) |
+| Tokens | 118.7M |
+| Batch size | 16 x 2 grad accum = 32 |
+| Sequence length | 512 |
+| LR | 3e-4 (cosine, warmup 100) |
+| Optimizer | AdamW (fused) |
+| Hardware | 1x Tesla T4 (fp16) |
+| Training time | ~61 min |
+| Avg train loss | 1.319 |
+
+Full metrics: [`report.json`](report.json), [`eval_history.json`](eval_history.json),
+[`eval_before.json`](eval_before.json).
 
 ## Pretraining corpus
 Mixed on-the-fly from flat token streams:
@@ -24,25 +44,13 @@ Mixed on-the-fly from flat token streams:
 
 Corpus built by [`prepare_domain_corpus.py`](https://github.com/Njap-png/Agron/blob/main/prepare_domain_corpus.py).
 
-## Training details
-| Field | Value |
-|---|---|
-| Init weights | teddyoas/Agron (model.safetensors) |
-| Steps | 5000 |
-| Batch size | 16 × 1 grad acc = 16 |
-| Sequence length | 512 |
-| LR | 3e-4 (cosine, warmup 200) |
-| Optimizer | AdamW (fused) |
-| Tokens seen | ~41.0M |
-| Hardware | 1× Tesla T4 (fp16) |
-| Training script | [`pretrain_agron_domain.py`](https://github.com/Njap-png/Agron/blob/main/pretrain_agron_domain.py) |
-
 ## Held-out evaluation
+
 | Source | Before (loss/PPL) | After (loss/PPL) | Δ |
 |---|---|---|---|
-| code | 6.740 / 845.3 | 1.453 / 4.28 | -5.287 |
-| cyber | 7.632 / 2063.7 | 1.358 / 3.89 | -6.274 |
-| replay (wikitext) | 7.208 / 1350.8 | 5.626 / 277.6 | -1.582 |
+| code | 1.453 / 4.28 | 1.245 / 3.47 | -0.208 |
+| cyber | 1.358 / 3.89 | 1.064 / 2.90 | -0.294 |
+| replay (wikitext) | 5.626 / 277.6 | 5.024 / 152.0 | -0.602 |
 
 ![Domain pretraining curves](domain_pretraining_curve.png)
 
@@ -50,12 +58,11 @@ Corpus built by [`prepare_domain_corpus.py`](https://github.com/Njap-png/Agron/b
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 tok = AutoTokenizer.from_pretrained("teddyoas/Agron")
-model = AutoModelForCausalLM.from_pretrained("teddyoas/Agron")  # this repo updated with these weights
+model = AutoModelForCausalLM.from_pretrained("teddyoas/Agron")  # promoted 1-epoch weights
 ```
-Or load from `teddyoas/Agron` directly (weights replaced by this upload).
 
 ## Files
 - `model.safetensors`, `config.json`, `generation_config.json`
 - `tokenizer.json`, `tokenizer_config.json`
 - `domain_pretraining_curve.png`
-- `eval_history.json`, `report.json`, `samples.json`
+- `eval_history.json`, `eval_before.json`, `report.json`, `samples.json`