Update README.md
Browse files
README.md
CHANGED
|
@@ -39,32 +39,30 @@ Fine-tuned on ≈ 2.3k ScienceBase “data → metadata” pairs to automate cre
|
|
| 39 |
Generate schema-compliant metadata text from a JSON/CSV representation of a ScienceBase item.
|
| 40 |
|
| 41 |
### Downstream Use
|
| 42 |
-
Integrate as a micro-service in data-repository pipelines
|
| 43 |
|
| 44 |
### Out-of-Scope
|
| 45 |
-
Open-ended content generation,
|
| 46 |
|
| 47 |
---
|
| 48 |
|
| 49 |
## Bias, Risks, and Limitations
|
| 50 |
* Domain-specific bias toward ScienceBase field names.
|
| 51 |
* Possible hallucination of fields when prompts are underspecified.
|
| 52 |
-
* Knowledge limited to training corpus and Jan 2025 Llama 3 cutoff.
|
| 53 |
|
| 54 |
-
**Recommendation:** keep a human curator in the loop and validate output against your schema.
|
| 55 |
|
| 56 |
---
|
| 57 |
|
| 58 |
## Training Details
|
| 59 |
|
| 60 |
### Training Data
|
| 61 |
-
* ~2.3k ScienceBase records with curated metadata.
|
| 62 |
-
|
| 63 |
|
| 64 |
### Training Procedure
|
| 65 |
| Hyper-parameter | Value |
|
| 66 |
|-----------------|-------|
|
| 67 |
-
| Max sequence length |
|
| 68 |
| Precision | fp16 / bf16 (auto) |
|
| 69 |
| Quantisation | 4-bit QLoRA (`load_in_4bit=True`) |
|
| 70 |
| LoRA rank / α | 16 / 16 |
|
|
@@ -79,8 +77,8 @@ Open-ended content generation, legal/medical decisions, or any application outsi
|
|
| 79 |
| Field | Value |
|
| 80 |
|-------|-------|
|
| 81 |
| GPU | 1 × NVIDIA A100 80 GB |
|
| 82 |
-
| Total training hours |
|
| 83 |
-
|
|
| 84 |
|
| 85 |
### Software Stack
|
| 86 |
| Package | Version |
|
|
@@ -103,21 +101,13 @@ Open-ended content generation, legal/medical decisions, or any application outsi
|
|
| 103 |
*Evaluation still in progress.*
|
| 104 |
|
| 105 |
|
| 106 |
-
---
|
| 107 |
-
|
| 108 |
-
## Environmental Impact
|
| 109 |
-
| Field | Value |
|
| 110 |
-
|-------|-------|
|
| 111 |
-
| Hardware | 1 × A100-80 GB |
|
| 112 |
-
| Hours | ~120 hours |
|
| 113 |
-
| Cloud/HPC provider | ARM Cumulus HPC |
|
| 114 |
|
| 115 |
---
|
| 116 |
|
| 117 |
## Technical Specifications
|
| 118 |
|
| 119 |
### Architecture & Objective
|
| 120 |
-
|
| 121 |
|
| 122 |
---
|
| 123 |
|
|
|
|
| 39 |
Generate schema-compliant metadata text from a JSON/CSV representation of a ScienceBase item.
|
| 40 |
|
| 41 |
### Downstream Use
|
| 42 |
+
Integrate as a micro-service in data-repository pipelines.
|
| 43 |
|
| 44 |
### Out-of-Scope
|
| 45 |
+
Open-ended content generation, or any application outside metadata curation.
|
| 46 |
|
| 47 |
---
|
| 48 |
|
| 49 |
## Bias, Risks, and Limitations
|
| 50 |
* Domain-specific bias toward ScienceBase field names.
|
| 51 |
* Possible hallucination of fields when prompts are underspecified.
|
|
|
|
| 52 |
|
|
|
|
| 53 |
|
| 54 |
---
|
| 55 |
|
| 56 |
## Training Details
|
| 57 |
|
| 58 |
### Training Data
|
| 59 |
+
* ~ 2.3k ScienceBase records with curated metadata.
|
| 60 |
+
|
| 61 |
|
| 62 |
### Training Procedure
|
| 63 |
| Hyper-parameter | Value |
|
| 64 |
|-----------------|-------|
|
| 65 |
+
| Max sequence length | 100 000 |
|
| 66 |
| Precision | fp16 / bf16 (auto) |
|
| 67 |
| Quantisation | 4-bit QLoRA (`load_in_4bit=True`) |
|
| 68 |
| LoRA rank / α | 16 / 16 |
|
|
|
|
| 77 |
| Field | Value |
|
| 78 |
|-------|-------|
|
| 79 |
| GPU | 1 × NVIDIA A100 80 GB |
|
| 80 |
+
| Total training hours | ~10 hours |
|
| 81 |
+
| Cloud/HPC provider | ARM Cumulus HPC |
|
| 82 |
|
| 83 |
### Software Stack
|
| 84 |
| Package | Version |
|
|
|
|
| 101 |
*Evaluation still in progress.*
|
| 102 |
|
| 103 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
|
| 105 |
---
|
| 106 |
|
| 107 |
## Technical Specifications
|
| 108 |
|
| 109 |
### Architecture & Objective
|
| 110 |
+
QLoRA-tuned `Llama-3.1-8B-Instruct`; causal-LM objective with structured-to-text instruction prompts.
|
| 111 |
|
| 112 |
---
|
| 113 |
|