PEFT
Safetensors
GGUF
English
llama
conversational
TravisPing commited on
Commit
f82da8e
·
verified ·
1 Parent(s): 1c242d1

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +8 -18
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; bootstrap metadata for legacy collections.
43
 
44
  ### Out-of-Scope
45
- Open-ended content generation, legal/medical decisions, 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
- * 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
- * Pre-processing: control-char stripping, field normalisation, incomplete rows removed.
63
 
64
  ### Training Procedure
65
  | Hyper-parameter | Value |
66
  |-----------------|-------|
67
- | Max sequence length | 20 000 |
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 | **TODO** |
83
- | Compute region / cluster | **TODO** |
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
- LoRA-tuned `Llama-3.1-8B`; causal-LM objective with structured-to-text instruction prompts.
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