Upload folder using huggingface_hub

1_Pooling/config.json

@@ -0,0 +1,10 @@
+{
+    "word_embedding_dimension": 384,
+    "pooling_mode_cls_token": false,
+    "pooling_mode_mean_tokens": true,
+    "pooling_mode_max_tokens": false,
+    "pooling_mode_mean_sqrt_len_tokens": false,
+    "pooling_mode_weightedmean_tokens": false,
+    "pooling_mode_lasttoken": false,
+    "include_prompt": true
+}

README.md

@@ -0,0 +1,266 @@
+---
+language: en
+license: apache-2.0
+library_name: sentence-transformers
+tags:
+- sentence-transformers
+- sentence-similarity
+- feature-extraction
+- resume-matching
+- job-matching
+- recruiting
+- talent-acquisition
+pipeline_tag: sentence-similarity
+base_model: sentence-transformers/all-MiniLM-L6-v2
+datasets:
+- custom
+model-index:
+- name: resumator
+  results: []
+---
+
+# Resumator
+
+A fine-tuned sentence-transformer model for **resume-to-job matching**. It encodes candidate resumes and job descriptions into a shared 384-dimensional embedding space, enabling fast semantic similarity search via cosine distance.
+
+## Model Details
+
+| Property | Value |
+|---|---|
+| **Base Model** | [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) |
+| **Architecture** | BERT (6 layers, 12 attention heads) |
+| **Embedding Dimensions** | 384 |
+| **Max Sequence Length** | 512 tokens |
+| **Pooling** | Mean token pooling + L2 normalization |
+| **Similarity Function** | Cosine similarity |
+| **Model Size** | ~91 MB (safetensors) |
+| **Training Loss** | CosineSimilarityLoss |
+| **Training Data** | 624 resume-job pairs with LLM-generated match scores |
+| **Parameters** | ~22.7M |
+
+## Use Case
+
+Built specifically for **recruiting/talent matching** pipelines:
+- Encode candidate resumes (name, skills, experience, location, full resume text)
+- Encode job postings (title, company, location, description, requirements)
+- Find best matches via cosine similarity (pgvector, FAISS, or in-memory)
+
+The model understands domain-specific relationships like:
+- "React developer" ↔ "Frontend Engineer" (skill-title alignment)
+- "3 years Python" ↔ "Senior Python Developer" (experience-level mapping)
+- Resume structure (summary, work history, skills sections)
+
+## Usage
+
+### sentence-transformers
+
+```python
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer("shankerram3/resumator")
+
+# Encode a candidate
+candidate_text = """
+Name: Jane Doe
+Skills: Python, React, PostgreSQL, AWS
+Experience: 4 years
+Location: San Francisco, CA
+Resume: Full-stack engineer with 4 years building scalable web applications...
+"""
+
+# Encode a job
+job_text = """
+Title: Senior Software Engineer
+Company: TechCorp
+Location: San Francisco, CA
+Description: Looking for an experienced full-stack engineer with Python and React...
+"""
+
+embeddings = model.encode([candidate_text, job_text])
+similarity = embeddings[0] @ embeddings[1]  # cosine similarity (vectors are L2-normalized)
+print(f"Match score: {similarity:.4f}")
+```
+
+### With pgvector (PostgreSQL)
+
+```sql
+-- Store embeddings in pgvector columns (384 dimensions)
+ALTER TABLE candidates ADD COLUMN embedding vector(384);
+ALTER TABLE jobs ADD COLUMN embedding vector(384);
+
+-- Find top matching jobs for a candidate
+SELECT j.title, 1 - (c.embedding <=> j.embedding) AS similarity
+FROM candidates c, jobs j
+WHERE c.id = 123
+ORDER BY c.embedding <=> j.embedding
+LIMIT 20;
+```
+
+### Transformers (without sentence-transformers)
+
+```python
+from transformers import AutoTokenizer, AutoModel
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("shankerram3/resumator")
+model = AutoModel.from_pretrained("shankerram3/resumator")
+
+def encode(text):
+    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512, padding=True)
+    with torch.no_grad():
+        outputs = model(**inputs)
+    # Mean pooling + L2 normalize
+    embeddings = outputs.last_hidden_state.mean(dim=1)
+    return torch.nn.functional.normalize(embeddings, p=2, dim=1)
+```
+
+## Training
+
+### Approach
+
+Fine-tuned from `all-MiniLM-L6-v2` using **CosineSimilarityLoss** on 624 curated resume-job pairs. Each pair was scored by an LLM (scoring relevance from 0.0 to 1.0), and the model was trained to reproduce those similarity scores in embedding space.
+
+### Training Data Format
+
+Each training example is a (resume_text, job_text, similarity_score) triple:
+- **resume_text**: Structured candidate profile (name, titles, skills, experience, location, full resume)
+- **job_text**: Structured job posting (title, company, industry, location, description)
+- **similarity_score**: Float 0.0–1.0 from LLM evaluation
+
+### Why Fine-Tuning?
+
+Generic sentence-transformers treat resumes and job descriptions as arbitrary text. Fine-tuning teaches the model:
+1. **Domain vocabulary**: "OPT", "H1B", "C2C" are visa types, not random acronyms
+2. **Structural alignment**: Match skills sections to requirements sections
+3. **Experience calibration**: "3 years Java" is closer to "mid-level Java developer" than "senior architect"
+4. **Recruiting context**: Company culture descriptions have lower weight than technical requirements
+
+## Benchmarks
+
+Evaluated on 50 candidates × 50 jobs = 2,500 pairs from anonymized recruiting data.
+
+### Similarity Score Distribution
+
+| Metric | Value |
+|---|---|
+| **Mean** | 0.5326 |
+| **Median** | 0.5399 |
+| **Std Dev** | 0.0917 |
+| **Min** | 0.1378 |
+| **Max** | 0.8341 |
+| **P10** | 0.4203 |
+| **P25** | 0.4815 |
+| **P75** | 0.5939 |
+| **P90** | 0.6376 |
+| **P95** | 0.6668 |
+
+### Score Histogram
+
+```
+0.0-0.1:                                    0 ( 0.0%)
+0.1-0.2:                                    4 ( 0.2%)
+0.2-0.3: #                                 45 ( 1.8%)
+0.3-0.4: #####                            146 ( 5.8%)
+0.4-0.5: ####################             601 (24.0%)
+0.5-0.6: ########################        1151 (46.0%)
+0.6-0.7: #################                506 (20.2%)
+0.7-0.8: #                                 41 ( 1.6%)
+0.8-0.9:                                    6 ( 0.2%)
+0.9-1.0:                                    0 ( 0.0%)
+```
+
+The distribution shows good **discrimination** — most pairs score in the 0.4–0.6 range (mediocre match), while strong matches (>0.7) are rare and meaningful. A UI/UX designer candidate correctly scores 0.77 against UI/UX design jobs vs 0.55 against data engineering roles.
+
+### Sample Matches
+
+**Candidate: UI/UX Designer (anonymized)**
+| Rank | Score | Job |
+|---|---|---|
+| #1 | 0.7770 | UI/UX Designer |
+| #2 | 0.7343 | Graphic Designer |
+| #3 | 0.7338 | UI/UX Designer |
+| #4 | 0.7336 | Apparel Graphic Designer |
+| #5 | 0.7282 | Graphic Designer (UI/UX) & Video Producer |
+
+### Inference Speed
+
+Measured on Apple M-series CPU (single thread):
+
+| Operation | Latency |
+|---|---|
+| **Model load** (cold start) | 450 ms |
+| **Single embedding** (mean) | 8.4 ms (P50) |
+| **Single embedding** (worst case) | 335 ms (P99, first call warmup) |
+| **Batch of 10** | 144 ms (14.4 ms/item) |
+| **Batch of 25** | 141 ms (5.6 ms/item) |
+| **Batch of 50** | 272 ms (5.4 ms/item) |
+| **50 candidates** (full encode) | 210 ms |
+| **50 jobs** (full encode) | 200 ms |
+
+Batch encoding is ~1.5x more efficient per item than single encoding.
+
+### Resource Usage
+
+| Metric | Value |
+|---|---|
+| Peak RAM (inference) | ~300 MB |
+| Model file size | 91 MB |
+| Dependencies | torch (~200 MB), sentence-transformers (~50 MB) |
+
+## Input Format
+
+For best results, structure your input text consistently:
+
+**Candidate format:**
+```
+Name: {name}
+Titles: {current_title}, {previous_titles}
+Skills: {skill1}, {skill2}, {skill3}
+Experience: {years} years
+Location: {city}, {state}
+Visa: {visa_status}
+Resume: {full_resume_text}
+```
+
+**Job format:**
+```
+Title: {job_title}
+Company: {company_name}
+Industry: {industry}
+Location: {city}, {state}
+Experience: {required_years} years
+Description: {full_job_description}
+```
+
+## Limitations
+
+- **English only** — trained exclusively on English resumes and US job postings
+- **512 token limit** — long resumes/descriptions are truncated; key info should be early in the text
+- **US tech market bias** — trained on US tech recruiting data; may not generalize well to other markets or industries
+- **Small training set** — 624 pairs; performance may vary on underrepresented roles or industries
+- **No temporal awareness** — doesn't account for job posting freshness or career progression timing
+
+## Model Architecture
+
+```
+SentenceTransformer(
+  (0): Transformer (BertModel, 6 layers, 384 hidden)
+  (1): Pooling (mean tokens)
+  (2): Normalize (L2)
+)
+```
+
+## Citation
+
+```bibtex
+@misc{resumator2026,
+  title={Resumator: Fine-tuned Sentence Transformer for Resume-Job Matching},
+  author={Shanker Ram},
+  year={2026},
+  url={https://huggingface.co/shankerram3/resumator}
+}
+```
+
+## License
+
+Apache 2.0

config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "BertModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 6,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "transformers_version": "4.57.1",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522
+}

config_sentence_transformers.json

@@ -0,0 +1,14 @@
+{
+  "__version__": {
+    "sentence_transformers": "5.1.2",
+    "transformers": "4.57.1",
+    "pytorch": "2.9.0"
+  },
+  "model_type": "SentenceTransformer",
+  "prompts": {
+    "query": "",
+    "document": ""
+  },
+  "default_prompt_name": null,
+  "similarity_fn_name": "cosine"
+}

eval/similarity_evaluation_results.csv

@@ -0,0 +1,4 @@
+epoch,steps,cosine_pearson,cosine_spearman
+1.0,20,0.8070555830833193,0.5252090240303582
+2.0,40,0.8526556431202823,0.6380171066139698
+3.0,60,0.8607002737646279,0.6508924200349745

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b4b34f5117868f069b5bdfed455952c99945c76f66bbe12cd1b865a997b4382
+size 90864192

modules.json

@@ -0,0 +1,20 @@
+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "sentence_transformers.models.Transformer"
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Pooling",
+    "type": "sentence_transformers.models.Pooling"
+  },
+  {
+    "idx": 2,
+    "name": "2",
+    "path": "2_Normalize",
+    "type": "sentence_transformers.models.Normalize"
+  }
+]

sentence_bert_config.json

@@ -0,0 +1,4 @@
+{
+    "max_seq_length": 256,
+    "do_lower_case": false
+}

special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "100": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "101": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "[CLS]",
+  "do_basic_tokenize": true,
+  "do_lower_case": true,
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "max_length": 128,
+  "model_max_length": 256,
+  "never_split": null,
+  "pad_to_multiple_of": null,
+  "pad_token": "[PAD]",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "sep_token": "[SEP]",
+  "stride": 0,
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "[UNK]"
+}