README.md

---
license: cc-by-sa-3.0
library_name: transformers
tags:
- supertrainer2000
- human-data
datasets:
- euclaise/TinyCoT
- euclaise/reddit-instruct
- sablo/oasst2_curated
- euclaise/SciCoT
metrics:
- accuracy
---



*Now with a training bug fixed!*



![image/png](https://cdn-uploads.huggingface.co/production/uploads/64137e2150358a805203cbac/DlTWku8gant1yx6NaxqJX.png)

Memphis-CoT is a finetune of [StableLM 3b 4e1t](stabilityai/stablelm-3b-4e1t) on [TinyCoT](https://huggingface.co/datasets/euclaise/TinyCoT), [SciCoT](https://huggingface.co/datasets/euclaise/SciCoT), along with [reddit-instruct](https://huggingface.co/datasets/euclaise/reddit-instruct) (subset to 5000 examples, excluding posts with brackets in the title) and a [curated](https://huggingface.co/datasets/sablo/oasst2_curated) subset of [oasst2](https://huggingface.co/datasets/OpenAssistant/oasst2).

**Memphis was trained *only* on human data! No GPT generations here.**

Finetuning was performed using my [supertrainer2000](https://github.com/euclaise/supertrainer2000) framework, using my Adalite optimizer.


## Training Procedure
I finetuned the model using an iterative rationale-bootstrapping procedure inspired by [STaR](https://research.google/pubs/star-self-taught-reasoner-bootstrapping-reasoning-with-reasoning/) and [SPIN](https://arxiv.org/abs/2401.01335)

First, I finetuned the model on all the datasets using a [MixCE](https://arxiv.org/abs/2305.16958) loss and [NEFTune](https://arxiv.org/abs/2310.05914), for 2 epochs.

I then performed the following steps 3 times:
1. Generate responses for each question in TinyCoT using the current model, check each response for correctness, and create a dataset of (correct, incorrect) pairs. Extra values are discarded, such that each correct and incorrect response is unique.
2. Finetune the model for 1 epoch using a ranking loss over length-normalized log-probabilities of each sequence, similar to [Preference Ranking Optimization](https://arxiv.org/abs/2306.17492), comparing the correct vs incorrect generated response. Additionally, a standard CE loss over the chosen completion was included.

This should be more efficient than either STaR or SPIN, as it uses a ranking loss rather than rejection sampling (unlike STaR), and verifies correctness instead of assuming all model responses are incorrect (unlike SPIN).

To prevent excessive drift, I kept the model weights as a moving average: After each generate+train cycle, I interpolated between the previous model weights and the updated weights using spherical linear interpolation (SLERP), with an interpolation factor of 0.99.

## Prompt formats

The format for reddit-instruct and oasst2 was:

```
### User:
[insert instruction here]
### Assistant:
[insert response here]
### User:
...
```

The format for TinyCoT was:
```
### User:
[insert instruction here]
### Rationale:
[insert reasoning here]
### Answer:
[insert direct answer here]
```

## Benchmarks

| Model                                                                  | Size   | Data                | Method        | GSM8K (5-shot) | AGIEval (English/Nous subset, acc_norm) | BIG Bench Hard (CoT, few-shot*) |
|:-----------------------------------------------------------------------|--------|:--------------------|---------------|:---------------|:----------------------------------------|:------------------------------  |
| [StableLM 3B Base](https://hf.co/stabilityai/stablelm-3b-4e1t)       | 3B     | Base                | Base          |    2.05%       | 25.14%                                  |  36.75%                           |
| [StableHermes 3B](https://hf.co/cxllin/StableHermes-3b)                | 3B     | GPT                 | SFT           |    3.64%       | 24.31%                                  | *37.28%*                        |
| [MPT 7B Instruct](https://hf.co/mosaicml/mpt-7b-instruct)              | **7B** | **Human**+Anthropic | SFT           |    2.05%       | 24.12%                                  | 11.01%                          |
| [OpenLLaMA 7B v2 open-instruct](http://hf.co/VMware/open-llama-7b-v2-open-instruct) | **7B** | **Human** (nearly: ecqa is an exception) | SFT | 8.64% | 23.21%                   | 29.84%                          |
| [StableLM Zephyr 3B](https://hf.co/stabilityai/stablelm-zephyr-3b)     | 3B     | GPT                 | DPO           |    possibly contaminated (45.72%)  | **33.31%**                   | 0.91%                           |
| [**Memphis-CoT 3B**](https://hf.co/euclaise/Memphis-CoT-3B)            | 3B     | **Human**           | Self-teaching |    **18.8%**       | *27.22%*                            | **36.92%**                      |

*5-shot, as performed automatically by LM Evaluation Harness bbh_cot_fewshot even with num_fewshot=0

Memphis outperforms other primarily-human-data models that are over twice its size, along with SFT models of its size, and trades with the Zephyr DPO model. That said, Zephyr uses synthetic data, and *much* more of it.

Note that BBH results have wide SEs, sometimes even exceeding 16%.


It is unclear why Zephyr performs so poorly on BBH. Perhaps it is overfit, or maybe there was an issue with vllm.

Notes:
- Evaluations were performed using the `agieval` branch of [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) (commit `0bef5c9c273b1c2f68e6018d4bb9c32b9aaff298`), using the `vllm` model.
- I tried to find human-data-trained StableLM models, but couldn't find any. I did find a few OpenLLaMA models, but they wouldn't load with LM Eval Harness and vllm. (I believe this can be fixed by changing the xformers backend, but I'm too lazy for that)
- OpenLLaMA 7B v2 open-instruct is a particularly relevant comparison, as it was trained on a *very* similar dataset.

## Hyperparameters

For the initial supervised finetuning step:
- Adalite optimizer, default hyperparameters of supertrainer2000 unless otherwise specified
- Lambda (Adalite's analogue to weight decay, see [here](https://arxiv.org/abs/2103.06583) for details) of 0.01
- LR of 1e-5
- MixCE ratio of 0.75
- Sequence length of 4096
- Cosine decay with a 20% warmup
- Frozen embeddings
- No training on inputs
- Accumulated batch size of 128
- NEFTune with an alpha of 10

For the generations:
- Generated using the current git version of `vllm`
- N=8
- Temperature of 0.5
- `top_p` of 0.8
- Maximum of 512 generated tokens, discarding responses that do not have a valid rationale and answer

For the rank finetuning:
- Adalite optimizer, default hyperparameters of supertrainer2000 unless otherwise specified
- Lambda of 0.01
- LR of 5e-7
- Rank loss weight of 0.25
- Sequence length of 1024
- Cosine schedule with 10% warmup
- Frozen embeddings
- No training on inputs
- Accumulated batch size of 128
- NEFTune with an alpha of 10