| --- |
| language: |
| - en |
| library_name: transformers |
| base_model: LiquidAI/LFM2.5-350M |
| tags: |
| - schema-extraction |
| - json-extraction |
| - structured-output |
| - information-extraction |
| - lora |
| - edge-ai |
| - schemalm |
| --- |
| |
| # SchemaLM v2 — LFM2.5 350M |
|
|
| A 350M parameter model fine-tuned for **JSON-schema-conditioned information extraction**. |
| Give it a JSON Schema and any unstructured text — it returns a JSON object that matches the schema exactly. |
|
|
| **Base model:** [`LiquidAI/LFM2.5-350M`](https://huggingface.co/LiquidAI/LFM2.5-350M) (SSM hybrid, ideal for edge/CPU deployment) |
|
|
| --- |
|
|
| ## Benchmark Results |
|
|
| | Metric | Score | |
| |---|---| |
| | Overall score (v2 training benchmark) | **97.7%** | |
| | JSON validity | **100%** | |
| | Schema conformance | **100%** | |
| | Field extraction accuracy | **97.6%** | |
| | Internal benchmark cases | **96.2%** | |
|
|
| Evaluated on a 25-case internal benchmark covering easy/medium/hard extraction across 7 domains. |
|
|
| | Domain | Avg Accuracy | |
| |---|---| |
| | Software Projects | ~74% | |
| | Governance / Compliance | ~75% | |
| | HR | ~85% | |
| | Education | ~81% | |
| | Sales | ~55% | |
| | Healthcare | ~53% | |
| | Support (phone) | ~61% | |
|
|
| --- |
|
|
| ## Training |
|
|
| - **Dataset:** 49K synthetic schema-extraction examples (93 schemas, 29 domains, nested object/array coverage) |
| - **Method:** PEFT/LoRA SFT — rank 32, alpha 64, ~2% trainable parameters |
| - **Epochs:** 5 | **LR:** 2e-4 (cosine) | **Max seq:** 768 tokens |
| - **Precision:** bf16 on Ampere/Hopper, fp16 otherwise |
|
|
| --- |
|
|
| ## Usage |
|
|
| The model follows the **ChatML format**: system message = JSON Schema, user message = unstructured text, assistant response = JSON. |
|
|
| ### Basic (Transformers) |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch, json |
| |
| model_id = "senthil090/schemalm-v2-lfm2-350m" |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, torch_dtype=torch.bfloat16, device_map="auto" |
| ) |
| model.eval() |
| |
| def extract(schema: dict, text: str, max_new_tokens: int = 256) -> dict: |
| messages = [ |
| {"role": "system", "content": json.dumps(schema)}, |
| {"role": "user", "content": text}, |
| ] |
| prompt = tokenizer.apply_chat_template( |
| messages, tokenize=False, add_generation_prompt=True |
| ) |
| inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
| with torch.no_grad(): |
| out = model.generate( |
| **inputs, |
| max_new_tokens=max_new_tokens, |
| do_sample=False, |
| repetition_penalty=1.08, |
| ) |
| raw = tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True) |
| return json.loads(raw.strip()) |
| ``` |
|
|
| --- |
|
|
| ### Example 1 — Sprint planning note → structured task |
|
|
| ```python |
| schema = { |
| "type": "object", |
| "properties": { |
| "task_title": {"type": "string"}, |
| "assignee": {"type": "string"}, |
| "priority": {"type": "string", "enum": ["critical", "high", "medium", "low"]}, |
| "story_points": {"type": "integer"}, |
| "status": {"type": "string", "enum": ["todo", "in_progress", "done"]}, |
| "epic": {"type": "string"}, |
| }, |
| "required": ["task_title", "assignee", "priority"], |
| } |
| |
| text = ( |
| "Assign the Redis caching implementation to Priya for Sprint 14. " |
| "It's high priority under the Performance epic, estimated 5 story points. " |
| "Mark it as in progress since she already started." |
| ) |
| |
| result = extract(schema, text) |
| # → { |
| # "task_title": "Redis caching implementation", |
| # "assignee": "Priya", |
| # "priority": "high", |
| # "story_points": 5, |
| # "status": "in_progress", |
| # "epic": "Performance" |
| # } |
| ``` |
|
|
| --- |
|
|
| ### Example 2 — Nested support ticket (nested objects + arrays) |
|
|
| ```python |
| schema = { |
| "type": "object", |
| "properties": { |
| "customer": { |
| "type": "object", |
| "properties": { |
| "name": {"type": "string"}, |
| "company": {"type": "string"}, |
| "email": {"type": "string", "format": "email"}, |
| }, |
| "required": ["name", "company"], |
| }, |
| "issue": { |
| "type": "object", |
| "properties": { |
| "summary": {"type": "string"}, |
| "severity": {"type": "string", "enum": ["low", "medium", "high", "critical"]}, |
| }, |
| "required": ["summary", "severity"], |
| }, |
| "actions": { |
| "type": "array", |
| "items": { |
| "type": "object", |
| "properties": { |
| "owner": {"type": "string"}, |
| "task": {"type": "string"}, |
| }, |
| "required": ["owner", "task"], |
| }, |
| }, |
| }, |
| "required": ["customer", "issue"], |
| } |
| |
| text = ( |
| "Support note: customer Priya Shah from Velocity Corp, email priya@velocity.io. " |
| "Dashboard charts won't load — severity high. " |
| "Action: Sarah to open an engineering ticket. Marcus to send ETA to customer." |
| ) |
| |
| result = extract(schema, text) |
| # → { |
| # "customer": {"name": "Priya Shah", "company": "Velocity Corp", "email": "priya@velocity.io"}, |
| # "issue": {"summary": "Dashboard charts won't load", "severity": "high"}, |
| # "actions": [ |
| # {"owner": "Sarah", "task": "open an engineering ticket"}, |
| # {"owner": "Marcus", "task": "send ETA to customer"} |
| # ] |
| # } |
| ``` |
|
|
| --- |
|
|
| ### Example 3 — Compliance finding with date and enum fields |
|
|
| ```python |
| schema = { |
| "type": "object", |
| "properties": { |
| "finding_id": {"type": "string"}, |
| "standard": {"type": "string", "enum": ["ISO27001", "SOC2", "HIPAA", "GDPR", "PCI-DSS"]}, |
| "clause": {"type": "string"}, |
| "severity": {"type": "string", "enum": ["critical", "major", "minor", "observation"]}, |
| "description": {"type": "string"}, |
| "due_date": {"type": "string", "format": "date"}, |
| "owner": {"type": "string"}, |
| "status": {"type": "string", "enum": ["open", "in_remediation", "closed"]}, |
| }, |
| "required": ["standard", "severity", "description", "due_date"], |
| } |
| |
| text = ( |
| "Audit finding AF-2024-089: ISO27001 clause A.9.4.2 violation — " |
| "privileged accounts are not protected with MFA. Severity: major. " |
| "Remediation owner: Chen, due by 2024-09-30. Status: in_remediation." |
| ) |
| |
| result = extract(schema, text) |
| # → { |
| # "finding_id": "AF-2024-089", |
| # "standard": "ISO27001", |
| # "clause": "A.9.4.2", |
| # "severity": "major", |
| # "description": "Privileged accounts are not protected with MFA", |
| # "due_date": "2024-09-30", |
| # "owner": "Chen", |
| # "status": "in_remediation" |
| # } |
| ``` |
|
|
| --- |
|
|
| ### Example 4 — CPU inference (no GPU) |
|
|
| ```python |
| model = AutoModelForCausalLM.from_pretrained( |
| "senthil090/schemalm-v2-lfm2-350m", |
| torch_dtype=torch.float32, # fp32 for CPU |
| ) |
| model.eval() |
| # Same extract() function as above — just slower (~5–15s per call on a modern CPU) |
| ``` |
|
|
| --- |
|
|
| ## Limitations |
|
|
| - Best on extraction tasks (reading text and pulling fields) — not a general-purpose chat model |
| - Weaker on very long schemas (>50 fields) or schemas with complex `pattern`/`format` constraints |
| - Sales numeric extraction and phone-call transcripts have lower accuracy (~50–60%) — targeted in v2.1 hardcase training |
| - Not suitable for code generation, reasoning, or creative tasks |
|
|
| ## Citation |
|
|
| ``` |
| @misc{schemalm-v2, |
| title = {SchemaLM v2 — JSON Schema-Conditioned Information Extraction}, |
| author = {Senthil}, |
| year = {2025}, |
| url = {https://huggingface.co/senthil090/schemalm-v2-lfm2-350m} |
| } |
| ``` |
