104 lines
3.9 KiB
Markdown
104 lines
3.9 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
library_name: transformers
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
base_model: thesreedath/slm-125m-base
|
|||
|
|
tags:
|
|||
|
|
- llama
|
|||
|
|
- legal
|
|||
|
|
- financial
|
|||
|
|
- small-language-model
|
|||
|
|
- sft
|
|||
|
|
- instruction-tuning
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# legal-slm-125m-sft
|
|||
|
|
|
|||
|
|
A 125M-parameter Llama-architecture model **supervised-fine-tuned to answer
|
|||
|
|
legal & financial questions**. It is the instruction-tuned sibling of a
|
|||
|
|
from-scratch base model: where the base model only *continues* text, this model
|
|||
|
|
*responds* to a question.
|
|||
|
|
|
|||
|
|
- **Base model:** [`thesreedath/slm-125m-base`](https://huggingface.co/thesreedath/slm-125m-base) (125M, pretrained 10 epochs)
|
|||
|
|
- **Fine-tuned on:** 5,846 grounded legal/financial Q&A pairs (synthetic, quality-filtered)
|
|||
|
|
- **Validation loss:** 2.06 (cross-entropy on held-out Q&A)
|
|||
|
|
|
|||
|
|
> **Base-capacity model.** At 125M parameters this is a study in end-to-end LLM
|
|||
|
|
> engineering, not a production assistant. It learns the *shape* of a good answer
|
|||
|
|
> and often gets the gist right, but it will **fabricate** case names, statutes,
|
|||
|
|
> and figures. **Never** use its output as legal, financial, or factual advice.
|
|||
|
|
|
|||
|
|
## What it does
|
|||
|
|
|
|||
|
|
The base model, given "In a breach of contract claim, the plaintiff…", would ramble
|
|||
|
|
onward. This model, asked a question, answers it:
|
|||
|
|
|
|||
|
|
> **Q:** In a breach of contract claim, what must the plaintiff prove?
|
|||
|
|
> **A:** The plaintiff must show: (1) a contract; (2) an agreement to perform a
|
|||
|
|
> specific act; (3) an obligation to perform…
|
|||
|
|
|
|||
|
|
## Chat format
|
|||
|
|
|
|||
|
|
The tokenizer has role special tokens but **no chat template**, so format prompts
|
|||
|
|
manually as `<|bos|><|system|>{system}<|user|>{question}<|assistant|>` and let the
|
|||
|
|
model complete the answer up to `<|eos|>`:
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
import torch
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
tok = AutoTokenizer.from_pretrained("jonam-ai/legal-slm-125m-sft")
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained("jonam-ai/legal-slm-125m-sft", torch_dtype=torch.bfloat16)
|
|||
|
|
|
|||
|
|
system = "You are a knowledgeable legal and financial assistant. Answer accurately and concisely."
|
|||
|
|
question = "What is the purpose of a Form 10-K filing?"
|
|||
|
|
|
|||
|
|
def sid(t): return tok.convert_tokens_to_ids(t)
|
|||
|
|
ids = (tok("<|bos|>", add_special_tokens=False)["input_ids"]
|
|||
|
|
+ [sid("<|system|>")] + tok(system, add_special_tokens=False)["input_ids"]
|
|||
|
|
+ [sid("<|user|>")] + tok(question, add_special_tokens=False)["input_ids"]
|
|||
|
|
+ [sid("<|assistant|>")])
|
|||
|
|
out = model.generate(torch.tensor([ids]), max_new_tokens=120, do_sample=True,
|
|||
|
|
temperature=0.7, top_p=0.9, eos_token_id=sid("<|eos|>"),
|
|||
|
|
pad_token_id=sid("<|pad|>"))
|
|||
|
|
print(tok.decode(out[0][len(ids):], skip_special_tokens=True))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Training data
|
|||
|
|
|
|||
|
|
5,846 grounded question–answer pairs, synthesized with a **teacher-LLM
|
|||
|
|
distillation** pipeline over a cleaned corpus of US case law, SEC filings, and
|
|||
|
|
educational web text:
|
|||
|
|
|
|||
|
|
1. **Chunk** the corpus into ~800-token passages.
|
|||
|
|
2. **Generate** (Gemini Flash-Lite): write self-contained Q&A answerable *only*
|
|||
|
|
from the passage, balanced across task types (QA, extraction, summarization,
|
|||
|
|
rewrite) and difficulty (easy → hard).
|
|||
|
|
3. **Judge** (Gemini Flash as LLM-as-judge): keep only pairs that are grounded,
|
|||
|
|
correct, and self-contained (score ≥ 4/5). ~78% pass.
|
|||
|
|
4. **Dedup** (exact + MinHash-LSH near-duplicate removal).
|
|||
|
|
5. **Format** as chat and **tokenize** with the base model's own tokenizer, with
|
|||
|
|
**loss computed only on the answer tokens**.
|
|||
|
|
|
|||
|
|
Mix: case-law 45% · SEC 45% · web 10%. Split: 5,554 train / 292 val.
|
|||
|
|
|
|||
|
|
## Training
|
|||
|
|
|
|||
|
|
| | |
|
|||
|
|
|---|---|
|
|||
|
|
| Method | full fine-tune (not LoRA) |
|
|||
|
|
| Hardware | 1 × NVIDIA L4 |
|
|||
|
|
| Epochs | 2 |
|
|||
|
|
| Tokens seen | ~1.0M |
|
|||
|
|
| LR | 3e-5, cosine decay, 3% warmup |
|
|||
|
|
| Precision | bf16 compute, fp32 master weights |
|
|||
|
|
| Time | ~80 seconds |
|
|||
|
|
|
|||
|
|
## Limitations
|
|||
|
|
|
|||
|
|
125M parameters, English only, 1,024-token context, not aligned/RLHF'd. It imitates
|
|||
|
|
the form of grounded answers learned from a synthetic dataset; factual reliability
|
|||
|
|
is limited by model size. Not legal or financial advice.
|