language:
- en
license: llama2
tags:
- generated_from_trainer
- finance
model-index:
- name: completed-model
results:
- task:
type: text-generation
dataset:
name: ai2_arc
type: ai2_arc
metrics:
- type: AI2 Reasoning Challenge (25-Shot)
value: 71.93
name: AI2 Reasoning Challenge (25-Shot)
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
name: Open LLM Leaderboard
- task:
type: text-generation
dataset:
name: hellaswag
type: hellaswag
metrics:
- type: HellaSwag (10-shot)
value: 86.82
name: HellaSwag (10-shot)
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
name: Open LLM Leaderboard
- task:
type: text-generation
dataset:
name: multiple
type: miltiple
metrics:
- type: MMLU (5-shot)
value: 70.38
name: MMLU (5-shot)
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
name: Open LLM Leaderboard
- task:
type: text-generation
dataset:
name: truthful_qa
type: truthful_qa
metrics:
- type: TruthfulQA (0-shot)
value: 65.21
name: TruthfulQA (0-shot)
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
name: Open LLM Leaderboard
- task:
type: text-generation
dataset:
name: winogrande
type: winogrande
metrics:
- type: Winogrande (5-shot)
value: 83.58
name: Winogrande (5-shot)
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
name: Open LLM Leaderboard
- task:
type: text-generation
dataset:
name: gsm8k
type: gsm8k
metrics:
- type: GSM8k (5-shot)
value: 61.79
name: GSM8k (5-shot)
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 71.93
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=gradientai/v-alpha-tross
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 86.82
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=gradientai/v-alpha-tross
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 70.38
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=gradientai/v-alpha-tross
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 65.21
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=gradientai/v-alpha-tross
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 83.58
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=gradientai/v-alpha-tross
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 61.79
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=gradientai/v-alpha-tross
name: Open LLM Leaderboard
Albatross is a collection of domain-specific language models for finance applications developed by Gradient.
This is the repository for an early, limited capability version, the v-alpha-tross, designed to showcase performance on
- mathematical reasoning
- tabular understanding
- open-book retrieval (RAG) & summarization
- conversational interface
Release versions of Albatross models are additionally trained on proprietary implementations of the latest architecture augmentation, expanded training and alignment data, and target reduced hallucination at retrieval, improved auditability, and multi-hop reasoning. To inquire for access to release versions, please reach out to contact@gradient.ai
Model description
The v-alpha-tross model is based on meta-llama/Llama-2-70b-hf, with additional, finance specific, pre-training, fine-tuning and instruction tuning.
This model substantially outperforms Llama2-70B models on H6 Average score, and GSM8K, with similar performance to mistralai/Mixtral-8x7B-Instruct-v0.1. It also reaches gpt-3.5-turbo performance in extracting information from tabular data like those found in SEC filings.
Intended use
The v-alpha-tross is intended as a demonstration of Gradient’s Albatross framework for developing large language models specific to the finance domain. We welcome additional research and development, but do not plan on continued internal development on this legacy model.
To get the expected performance, follow formatting requirements of Llama-2 chat, including INST and <<SYS>> tags, and <s> tokens.
Training Strategy
The Albatross framework overcomes deficiencies in general-purpose language models that arise in the face of solving tasks in the finance domain.
Release versions of Albatross use an expanded data universe for extended capabilities.
Pre-Training
A base Llama2-70B is further pre-trained on finance data since LLMs are poor at answering questions when their internal relevant document store is sparse [1].
To curate quality training data with low operational overhead we demo a novel data gathering approach:
- Crawl public repositories of text data. For
v-alpha-tross, we limited to Red Pajamas and https://github.com/. - Programmatically filter the crawled corpus to datasets not likely to be in the base model's training already, using a likelihood ratio test adapted from LiRA membership inference.[2]
- Human finance professionals review the (much smaller) filtered corpus to further remove low quality results.
[1] Kandpal, Nikhil, et al. "Large language models struggle to learn long-tail knowledge." International Conference on Machine Learning. PMLR, 2023.
[2] Carlini, Nicholas, et al. "Membership inference attacks from first principles." 2022 IEEE Symposium on Security and Privacy (SP). IEEE, 2022.
Fine-Tuning
Supervised fine-tuning (SFT) and direct preference optimization (DPO)[3] further enhances performance and alignment on finance-related tasks.
v-alpha-tross includes a subset of Albatross tuning goals: financial anchoring, mathematical reasoning, tabular understanding, conversational communication, summarization.
| Category | # Tokens (1Ms) | % of Total |
|---|---|---|
| Chat (e.g. ultrachat_200k) | 640 | 45.2 |
| Alignment * (e.g. orca_dpo) | 331 | 23.4 |
| Math * (e.g. Goat[4]) | 300 | 21.2 |
| Tabular * | 68 | 4.8 |
| Summarization (e.g. legal_summarization) | 52 | 3.7 |
| Open-book (e.g. selfrag) | 25 | 1.8 |
(*) = Proprietary or includes proprietary data sets
[3] Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D. and Finn, C., 2023. Direct preference optimization: Your language model is secretly a reward model. NeurIPS.
[4] Liu, Tiedong, and Bryan Kian Hsiang Low. "Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic Tasks." arXiv preprint arXiv:2305.14201 (2023).
Benchmarks
From a Llama-2-70B base, v-alpha-tross improves H6 metrics, and in particular GSM8k (arithmetic reasoning), scoring similar to Mixtral-8x7B-Instruct-v0.1. Relative to a subset of Open LLM Leaderboard [4] models which also use Llama-2-70B as a base, the model achieves state of the art results for the Average H6 score.
On financial table understanding (our new metric) the model is on par with GPT-3.5.
| Model | H6 [4] | GSM8k | sec_tables_v1 |
|---|---|---|---|
| v-alpha-tross | 73.28 | 61.79 | 100.0 |
| meta-llama/Llama-2-70B-hf | 67.87 | 54.06 | 75.76 |
| meta-llama/Llama-2-70b-chat-hf | 62.40 | 26.69 | 87.88 |
| mistralai/Mixtral-8x7B-Instruct-v0.1 | 72.70 | 61.11 | 82.35 |
| GPT-3.5 | N/A | 57.1 [5] | 100.0 |
[4] https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
[5] https://paperswithcode.com/sota/arithmetic-reasoning-on-gsm8k
Training procedure
We develop Albatross on Gradient’s distributed training platform, leveraging leading open source toolsets and optimizations like Axolotl, Accelerate, and Deepspeed for high throughput and memory efficiency.
Training hyperparameters (DPO)
The following hyperparameters were used during DPO training:
- learning_rate: 5e-07
- train_batch_size: 3
- eval_batch_size: 3
- seed: 42
- distributed_type: multi-GPU
- num_devices: 40
- total_train_batch_size: 120
- total_eval_batch_size: 120
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 18
- num_epochs: 1
- dpo_beta: .1
Framework versions
- Transformers 4.35.1
- Pytorch 2.0.1+cu118
- Datasets 2.14.7
- Tokenizers 0.14.1
Bias
v-alpha-tross has not been specifically aligned for safety, so the model can produce problematic outputs (especially when prompted to do so). It is also subject to any risks of the corpus that was used to train the base Llama 2 models.
More information & how to cite
Whitepaper coming soon!
The Gradient AI Team
Gradient is accelerating AI transformation across industries. https://gradient.ai/
Contact Us
Drop an email to contact@gradient.ai
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 73.28 |
| AI2 Reasoning Challenge (25-Shot) | 71.93 |
| HellaSwag (10-Shot) | 86.82 |
| MMLU (5-Shot) | 70.38 |
| TruthfulQA (0-shot) | 65.21 |
| Winogrande (5-shot) | 83.58 |
| GSM8k (5-shot) | 61.79 |