71 lines
2.8 KiB
Markdown
71 lines
2.8 KiB
Markdown
---
|
|
base_model: nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
|
|
base_model_relation: finetune
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
license: other
|
|
license_name: nvidia-nemotron-open-model-license
|
|
license_link: >-
|
|
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/
|
|
language:
|
|
- en
|
|
tags:
|
|
- figment
|
|
- v4
|
|
- merged-bf16
|
|
- protocol-navigation
|
|
- synthetic-data
|
|
- not-for-clinical-use
|
|
---
|
|
|
|
# Figment SFT V4 Merged BF16
|
|
|
|
This directory contains the `figment_sft_v4` LoRA adapter merged into `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16` with `peft.merge_and_unload(safe_merge=True)`.
|
|
|
|
## Artifact Summary
|
|
|
|
- Source adapter: `/checkpoints/figment_sft_v4/figment-sft-v4-lora`
|
|
- Merged output: `/checkpoints/figment_sft_v4/figment-sft-v4-lora-merged-bf16`
|
|
- Merge dtype: BF16
|
|
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
|
|
- Merged manifest SHA-256: `6678c0ec3a28817dba22eb9e7c682b9961f04bbfc688d0f1bcd137afaf8c8c38`
|
|
- HF shard 1 SHA-256: `1d95889e945363adcd70a0be54bc29407d49e28bf7a2c0415e1732d81d64186c`
|
|
- HF shard 2 SHA-256: `2a2e27563e78981c130349feece291c976cf7d5384690c6327795eef6d08d4c0`
|
|
- Matching GGUF file: `../v4-20260611-merged-bf16.gguf`
|
|
- GGUF SHA-256: `7e11f2295b101e9312f97075b8e48cabd8cc89539e92c8fa4218c4973aa31d8d`
|
|
|
|
## Training Data
|
|
|
|
The adapter was trained on the synthetic/de-identified `figment_sft_v4` corpus:
|
|
|
|
- Total rows: 1650
|
|
- Train rows: 1482
|
|
- Validation rows: 168
|
|
- Navigator-full rows: 1500
|
|
- Focused-repair rows: 150
|
|
- Full corpus SHA-256: `ef7a7c9a6a99927ba72ce244e03a9da3ab86d3cf5dc70786703fb5f8bdf2a289`
|
|
- Train split SHA-256: `f869d79da9ef670bc6479f8321e51b1f48cb5a16423265f34893a08e7648676e`
|
|
- Validation split SHA-256: `3ff7668b8216d6fa0be770d6d9ed5f1a0b12965f9312d5210b510807538738d3`
|
|
|
|
The corpus is published in the dataset repository `build-small-hackathon/figment-eval-traces` as the `figment_sft_v4` config.
|
|
|
|
## Evaluation
|
|
|
|
The main local field-holdout evaluation run was `local_4b_finetuned_v4_field_holdout_20260611T011930Z`:
|
|
|
|
| Metric | Value |
|
|
| --- | ---: |
|
|
| Total cases | 150 |
|
|
| Competence successes | 109/150 |
|
|
| Raw configured-model successes | 109/150 |
|
|
| Expected-label successes | 149/150 |
|
|
| Full fallback uses | 2 |
|
|
| Final validation successes | 148/150 |
|
|
| Model-visible fields retained | 1846/1950 |
|
|
|
|
The separate 50-case local evidence run `local_4b_finetuned_v4_evidence_20260611T0010Z` reached `37/50` competence successes, `37/50` raw configured-model successes, `0` full fallback uses, `50/50` final validation successes, and `624/650` model-visible fields retained.
|
|
|
|
## Safety
|
|
|
|
Prototype artifact only. This checkpoint is not a medical device and must not be used for diagnosis, prescribing, medication dosing, autonomous triage, or replacing local protocol, clinician judgment, supervisor review, or trained responder review.
|