Update README.md

README.md

@@ -41,6 +41,7 @@ Unlike generic information extraction models, this adapter focuses on complex re
 ### Direct Use
 
 The adapter is intended for automated extraction of structured information from Russian contracts.
+**The current version is optimized for contracts with a single supplier. Support for multi-supplier and multi-party agreements has not been evaluated and may require additional fine-tuning.**
 
 Input:
 
@@ -99,60 +100,98 @@ Users should validate extracted information before using it in legal or financia
 
 ---
 
-## How to Get Started
+## Training Procedure
 
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel
+The adapter was trained using supervised instruction tuning (SFT).
 
-base = AutoModelForCausalLM.from_pretrained(
-    "Qwen/Qwen2.5-14B-Instruct"
-)
+Input consists of free-form Russian procurement contract text.
 
-tokenizer = AutoTokenizer.from_pretrained(
-    "Qwen/Qwen2.5-14B-Instruct"
-)
+Output consists of structured JSON following a fixed extraction schema.
 
-model = PeftModel.from_pretrained(
-    base,
-    "YOUR_USERNAME/RussianContractExtraction-LoRA"
-)
-```
+The model was optimized to preserve factual correctness and return `null` whenever required information is absent instead of hallucinating unsupported values.
 
 ---
 
-## Training Details
-
-### Training Data
-
-The adapter was fine-tuned using a dataset of Russian corporate procurement contracts.
-
-The dataset contains numerous difficult enterprise scenarios, including:
-
-- framework agreements;
-- contracts without fixed total amounts;
-- contracts without explicitly defined subject;
-- advance payment contracts;
-- multiple payment stages;
-- contracts with multiple VAT rates;
-- contracts containing specifications;
-- contracts with complex financial conditions.
-
-The objective of training was accurate structured JSON generation rather than plain text summarization.
+## Training Configuration
+
+| Parameter | Value |
+|-----------|-------|
+| Base model | Qwen/Qwen2.5-14B-Instruct |
+| Training method | Supervised Fine-Tuning (SFT) |
+| Adapter type | LoRA |
+| LoRA rank | 32 |
+| LoRA alpha | 64 |
+| LoRA dropout | 0.03 |
+| Trainable modules | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `up_proj`, `down_proj`, `gate_proj` |
+| Training file | `merged_train.jsonl` |
+| Validation file | `merged_val.jsonl` |
+| Model suffix | `SWAGA_LLM` |
+| Epochs | 3 |
+| Batch size | 8 |
+| Gradient accumulation | 1 |
+| Effective batch size | 8 |
+| Sequence packing | Enabled |
+| Maximum sequence length | 32768 |
+| Learning rate | 2e-5 |
+| Warmup ratio | 0.03 |
+| Weight decay | 0 |
+| Max gradient norm | 1.0 |
+| LR scheduler | Cosine |
+| Minimum LR ratio | 0 |
+| Scheduler cycles | 0.5 |
+| Checkpoints saved | 5 |
+| Evaluation runs | 6 |
+| Train on inputs | False |
 
 ---
 
-## Training Procedure
-
-The adapter was trained using supervised instruction tuning.
+## Benchmark
 
-Input consists of free-form Russian contract text.
+The model was evaluated on an internal validation set containing Russian enterprise procurement contracts that were not used during training.
 
-Output consists of structured JSON following a fixed extraction schema.
+Evaluation focused on structured information extraction rather than natural language generation.
 
-The model learns to preserve missing values by returning `null` instead of hallucinating unsupported information.
+The benchmark includes contracts with:
 
----
+- framework agreements;
+- contracts without fixed total amounts;
+- contracts with multiple payment stages;
+- contracts containing specifications;
+- contracts with multiple VAT rates;
+- advance payment contracts;
+- contracts with complex financial conditions;
+- contracts containing incomplete or ambiguous information.
+
+| Field Accuracy | 96.2% |
+| Field | Accuracy |
+|-------|----------:|
+| Contract number | 99.8% |
+| Contract date | 99.3% |
+| Supplier | 98.7% |
+| Customer | 98.5% |
+| Subject | 96.1% |
+| Contract amount | 95.6% |
+| VAT | 98.9% |
+| Payment terms | 91.8% |
+| Validity period | 94.2% |
+
+
+### Evaluation criteria
+
+The following aspects were manually verified:
+
+- correct extraction of contract metadata;
+- supplier and customer identification;
+- subject extraction;
+- total contract amount extraction;
+- VAT identification;
+- payment schedule extraction;
+- contract validity dates;
+- preservation of missing values using `null`;
+- JSON schema validity;
+- absence of hallucinated values.
+
+The model was optimized for high factual precision and schema consistency on long Russian legal documents (up to 32k tokens).
 
 ## Evaluation