Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# tocsa/atlas-sft
|
| 2 |
+
|
| 3 |
+
**Atlas customer-support triage (capstone, Section 5)** — SFT model.
|
| 4 |
+
|
| 5 |
+
Atlas is a fictional SaaS project-management tool. The model reads an incoming support ticket and returns a `category`, an `urgency`, and a short drafted response — the Meridian pattern (eval -> SDG -> SFT -> DPO) applied end-to-end to a new domain during the capstone.
|
| 6 |
+
|
| 7 |
+
## Base model / architecture
|
| 8 |
+
|
| 9 |
+
nvidia/Mistral-NeMo-Minitron-8B-Instruct (MistralForCausalLM)
|
| 10 |
+
|
| 11 |
+
## Pipeline position
|
| 12 |
+
|
| 13 |
+
SFT model: a LoRA adapter trained on synthetically generated data, merged into the base model. This is the checkpoint used as the DPO starting point.
|
| 14 |
+
|
| 15 |
+
## Files
|
| 16 |
+
|
| 17 |
+
- `config.json` (702.0B)
|
| 18 |
+
- `generation_config.json` (111.0B)
|
| 19 |
+
- `model.safetensors` (15.7GB)
|
| 20 |
+
- `special_tokens_map.json` (414.0B)
|
| 21 |
+
- `tokenizer.json` (8.8MB)
|
| 22 |
+
- `tokenizer_config.json` (173.4KB)
|
| 23 |
+
|
| 24 |
+
## Recorded metrics
|
| 25 |
+
|
| 26 |
+
### baseline_eval.json
|
| 27 |
+
- **n**: 12
|
| 28 |
+
|
| 29 |
+
### best_sft.json
|
| 30 |
+
- **adapter_path**: /scratch/capstone_sft/checkpoints/epoch_2_step_56/model
|
| 31 |
+
- **epoch**: 2
|
| 32 |
+
|
| 33 |
+
## How to load
|
| 34 |
+
|
| 35 |
+
```python
|
| 36 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 37 |
+
|
| 38 |
+
model = AutoModelForCausalLM.from_pretrained("tocsa/atlas-sft")
|
| 39 |
+
tokenizer = AutoTokenizer.from_pretrained("tocsa/atlas-sft")
|
| 40 |
+
```
|
| 41 |
+
|
| 42 |
+
## Provenance
|
| 43 |
+
|
| 44 |
+
Produced in an NVIDIA instructor-led "Model Customization" workshop. Training data for this run was synthetically generated (NeMo Data Designer) and/or curated during the workshop exercises. Educational artifact — not reviewed for production use.
|