Developed by @BobaZooba | CV | LinkedIn | [email protected]
Model based on Mistral-7B-v0.1
GitHub Repo | Detailed step-by-step guide how to train this model
| HuggingFace Hub | 7B | 7B-GPTQ |
|---|---|---|
| Shurale-v1 | Link | Link |
- Shurale is an open-domain dialogue model for chit-chat conversations
- The model has the capability to establish a character and situation in the conversation
- It's a 7B model based on Mistral-7B-v0.1
- The model was trained using 1,112,000 dialogs for 10,000 steps with a batch size of 128
- Trained on 334 million tokens
- Maximum length at training was 2048 tokens
- The total cost of training this model is just $58
Shurale [/สสrษหlสฒe/] is a forest spirit in Bashkir and Tatar mythology.
Do you want models as cool as this one?
The parts of the dialogue (narrative and phrases) are separated using a newline symbol: \n
The maximum length during training was 2048 tokens. The SODA dataset was used for the training process.
The model in the prompt needs two things from you: narrative (description of the dialog) + dialog (dialog).
The narrative outlines the context for the conversation. This can include details about the situation, characters, and what's happening. This usually takes the form of 2-3 sentences.
The dialog is a series of phrases or lines from individuals involved. You just need to indicate who is saying what.
Details
Training examples consisted of both the narrative and the dialogue itself, with the participants' names clearly indicated.
| Narrative | A description of the situation within the dialogue |
| Characters names | A list of names of the characters participating in the dialogue |
| Phrases | Phrases used by the participants in the dialogue |
Narratives were deliberately omitted from 5% of the training examples, allowing the model to maintain a dialogue even without a narrative. However, using the model without a narrative is generally not recommended.
The baton was passed to Garry who then became the boss. He ran the show with an iron fist, making sure that everything
was done his way. No one dared to cross him for fear of being on the receiving end of his wrath
Garry: What the hell is going on around here? I thought I told you to get this place in order!
Bob:
In this example, we explicitly tell the model that it's now Bob's turn to speak. The end of the reply can be designated either by a newline symbol or by the name of the first character followed by a colon (Garry:).
I'm sorry, Garry. We're just having a little trouble getting everything done the way you want it.
Garry: Well, you need to fix it, and fast! I'm not going to tolerate any more slacking off! Do you understand me?
Bob: Yes, Garry. I understand.
Garry: Good. Now get to work and fix this mess!
Bob: Yes, Garry. I'll get right on it.
Garry: And don't give me any lip about it either! I'm not putting up with that kind of attitude around here any
more! Do you hear me?
Bob: Yes, Garry. I hear you loud and clear. I'll get to work right away and make sure everything is done your way.
Garry: Good. I'll be checking on you regularly to make sure that this is all getting done the way I want it! So
don't you dare let me down!
Recommended generation parameters for sampling:
| Param | Value |
|---|---|
| top_p | 0.75 |
| typical_p | 0.95 |
| top_k | 50 |
| temperature | 0.75 |
| repetition_penalty | 1.05 |
- Load model
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("BobaZooba/Shurale7B-v1")
model = AutoModelForCausalLM.from_pretrained("BobaZooba/Shurale7B-v1")- Run generation
input_text = "Dialog between two colleagues: Emma and Anna.\nEmma:"
tokenized = tokenizer(
input_text,
return_tensors="pt"
).to("cuda:0")
generated_indices = model.generate(
**tokenized,
do_sample=True,
max_new_tokens=128,
top_p=0.9
)[0].cpu()
print(tokenizer.decode(generated_indices))Run model as a service using HuggingFace ๐ค inference server: https://github.com/huggingface/text-generation-inference#get-started
1. Start a docker container with the model
model=BobaZooba/Shurale7B-v1
volume=$PWD/data
version=1.1.0 # please make sure you are using latest or stable version (>= 1.1.0)
docker run --gpus all --shm-size 1g -p 8081:80 -v \
$volume:/data ghcr.io/huggingface/text-generation-inference:$version \
--model-id $model --max-batch-prefill-tokens 2048 --dtype bfloat16If you want to run a model at RunPod you can find ready to use template by name "Shurale7B-v1" at RunPod. Please note that port 8081 is used to run this template.
https://www.runpod.io/console/gpu-cloud
| Field | Value |
|---|---|
| Container Image | ghcr.io/huggingface/text-generation-inference:1.1.0 |
| Docker Command | --model-id BobaZooba/Shurale7B-v1 --num-shard 1 --port 8081 --max-batch-prefill-tokens 2048 --dtype bfloat16 --json-output |
| Container Disk | 5 |
| Volume Disk | 15 |
| Volume Mount Path | /data |
| Expose HTTP Ports | 8081,8080 |
| Expose TCP Ports | 8082 |
2. Send request to the server and parse the response
import requests
import json
url = "127.0.0.1:8081/generate"
headers = {"Content-Type": "application/json"}
data = {
"inputs": "Dialog between two colleagues: Emma and Anna.\nEmma:",
"parameters": {
"max_new_tokens": 128,
"do_sample": True,
"top_p": 0.9,
"stop": ["\n"]
}
}
response = requests.post(url, headers=headers, data=json.dumps(data))
print(response.json()["generated_text"].strip())
# Hello, Anna! How was your evening?Or
pip install text-generationfrom text_generation import Client
input_text = "Dialog between two colleagues: Emma and Anna.\nEmma:"
client = Client("http://127.0.0.1:8081")
print(client.generate(input_text, max_new_tokens=128).generated_text)
text = ""
for response in client.generate_stream(input_text, max_new_tokens=20):
if not response.token.special:
text += response.token.text
print(text)The model was trained using only the training part of the SODA dataset.
This model, based on Mistral-7B-v0.1, was trained on over 1.1 million dialogues using 8 RTX 3090 (24 Gb) GPUs. The training process lasted 45 hours and made use of advanced techniques such as QLoRA (int4), DeepSpeed Stage 2, and gradient checkpointing. Flash Attention 2 was disabled due to this technique was not implemented for the model Mistral-7B-v0.1 at the moment of training.
| Field | Value |
|---|---|
| Model | Mistral-7B-v0.1 |
| Training steps | 10,000 |
| Warm up steps | 1,000 |
| Num epochs | 1.14 |
| Num training samples | 1,119,582 dialogs |
| Max sequence length | 2048 tokens |
| Num training tokens per epoch | 292,851,543 |
| Num training tokens total | 334,812,435 |
| Batch size | 4 |
| Gradient accumulation steps | 4 |
| GPUs | 8 x RTX 3090 (24 Gb) |
| Global batch size | 128 |
| Max batch tokens | 262,144 |
| Loss | 1.93 |
| Perplexity | 6.9 |
| Cost | $58 |
| Price per hour | $2.13 |
| Training time | 27 hours |
| Provider | vast.ai |
| Field | Value |
|---|---|
| Use gradient checkpointing | True |
| Use bnb int4 | True |
| Apply LoRA | True |
| LoRA rank | 64 |
| LoRA alpha | 32 |
| LoRA layers | all |
| Scheduler | WarmupDecayLR |
| Max lr | 2e-4 |
| Use Flash Attention 2 | False (not supported yet for mistal models |
| DeepSpeed Stage | 2 |
| DeepSpeed Offloading | True |
Detailed config
| Field | Value |
|---|---|
| save_safetensors | True |
| use_gradient_checkpointing | True |
| trainer_key | lm |
| force_fp16 | False |
| from_gptq | False |
| deepspeed_stage | 2 |
| fsdp_strategy | |
| seed | 42 |
| stabilize | True |
| Field | Value |
|---|---|
| dataset_key | soda |
| train_local_path_to_data | ./train.jsonl |
| eval_local_path_to_data | None |
| shuffle | True |
| Field | Value |
|---|---|
| tokenizer_name_or_path | None |
| tokenizer_use_fast | None |
| tokenizer_padding_side | None |
| Field | Value |
|---|---|
| collator_key | lm |
| max_length | 2048 |
| Field | Value |
|---|---|
| model_name_or_path | mistralai/Mistral-7B-v0.1 |
| model_type | llama |
| use_flash_attention_2 | True |
| trust_remote_code | True |
| device_map | None |
| Field | Value |
|---|---|
| model_name_or_pathload_in_8bit | False |
| load_in_4bit | True |
| llm_int8_threshold | 6.0 |
| llm_int8_has_fp16_weight | True |
| bnb_4bit_use_double_quant | True |
| bnb_4bit_quant_type | nf4 |
| Field | Value |
|---|---|
| output_dir | ./outputs/ |
| per_device_train_batch_size | 4 |
| gradient_accumulation_steps | 4 |
| warmup_steps | 1000 |
| max_steps | None |
| num_train_epochs | 1 |
| learning_rate | 2e-4 |
| max_grad_norm | 1.0 |
| weight_decay | 0.001 |
| label_smoothing_factor | 0.1 |
| logging_steps | 10 |
| save_steps | 100 |
| save_total_limit | 1 |
| push_to_hub | True |
| Field | Value |
|---|---|
| report_to_wandb | True |
| Field | Value |
|---|---|
| apply_lora | True |
| lora_rank | 64 |
| lora_alpha | 32 |
| lora_dropout | 0.1 |
| lora_target_modules | all |
The model was trained on a synthetic dataset generated using ChatGPT, leading to a few critical issues with the current version. Often, the model tends to be rather bland and can occasionally be unnatural. Conversations can be very short, the model tends to say goodbye. Although the model wasn't explicitly trained to be safe, it's likely these traits are inherited from ChatGPT. Moreover, handling very long dialogues is considered out-of-domain for the model since it was trained with a maximum length of 2048 tokens. The model's ability to generate truth-valid facts wasn't tested, but it's probable that its performance in this area lags behind OpenAI models. Also, this model wasn't explicitly trained to follow instructions.
It is suggested to set a maximum context length, for example, 10 messages. Then, store the context in some form of data storage, such as a database. It is recommended to feed the model with the narrative and the last 10 messages. This way, the model will consistently receive the last 10 dialogue messages at each generation step.
def generate(prompt: str) -> str:
...
max_context_length = 10
narrative = "..."
separator = "\n"
bot_prompt = "Bot"
user_prompt = "Person"
context = list()
while True:
user_phrase = input("You: ")
context.append(f"{user_prompt}: {user_phrase}")
model_prompt = separator.join(
[narrative] + context[-max_context_length:] + [f"{bot_prompt}:"]
)
generated_response = generate(model_prompt)
bot_phrase = f"{bot_prompt}: {generated_response}"
context.append(bot_phrase)
print(bot_phrase)Tale Quest is my personal project which was built using xllm and Shurale. It's an interactive text-based game
in Telegram with dynamic AI characters, offering infinite scenarios
You will get into exciting journeys and complete fascinating quests. Chat
with George Orwell, Tech Entrepreneur, Young Wizard, Noir Detective, Femme Fatale and many more
Try it now: https://t.me/talequestbot
Default examples (not as interesting as in TaleQuest):
Example #1
Narrative: Mckenzi felt as though he had made a mistake as soon as he handed in his notice. His boss had been angry, and his co-workers disappointed. Mckenzi himself felt regretful and unsure of what the future held.
Dialog: Mckenzi: I'm sorry if this comes as a surprise, but I've decided to quit. Boss: What? Why? Is there something wrong? Mckenzi: No, nothing's wrong. I just feel like it's time for a change. Boss:
Original response: I see. Well, I'm sorry to hear that you're leaving. We'll miss you around here.
Shurale responses:
- I see. Well, we're going to miss you around here. You're a good worker and we're not used to seeing talented people leave us.
- I see. Well, we're going to miss you around here. You're a great employee and we'll be sad to see you go. Do you have any plans for what you'll do next?
- Well, I'm sorry to hear that. We're going to miss you around here. But if you're sure this is what you want to do, then we won't stand in your way.
- I don't understand. You've only been with us for a couple of months. Are you sure this is what you want to do?
- A change? Mckenzi, you've been with this company for years. You're one of our most valued employees. What could possibly make you want to leave?
Example #2
Narrative: Kanon felt her face flush with anger. She balled up her fists, digging her nails into her palms as she glared at Suzette. It was all she could do to keep herself from launching across the table and strangling him.
Dialog: Kanon: What the hell do you mean by that? Do you think I'm some kind of idiot? Suzette: No, of course not. I just think that you might not be aware of what you're getting yourself into. This whole thing could be a lot more complicated than you realize. Kanon:
Original response: I'm perfectly aware of what I'm doing, thank you. And I can handle myself just fine. So don't try to tell me what to do.
Shurale responses:
- Oh, please. Like you're some big expert on the situation or something? You don't know anything about me or my family or what we're going through. So why don't you just back off and let me handle this my way!
- I'm not an idiot, and I'm fully aware of what I'm getting myself into. I know exactly what I'm doing, and I'm not going to let you stop me.
- C'mon, Suzette. You know me better than that. I can handle my own business. Besides, it's not like I'm going to get involved with him on any level other than professional.
- I know exactly what I'm doing. I'm not some stupid little girl who gets caught up in things without knowing what she's doing.
- I can handle whatever comes my way. I'm not going to let some stupid rumors stop me from finding out the truth.
Example #3
Narrative: Kelcey is irresponsible. She often fails to finish articles she starts writing, and this has led to her being fired from several jobs.
Context: Kelcey:
Original response: Hey, boss. I'm sorry I didn't finish that article you asked me to write.
Shurale responses:
- I'm sorry, boss. I know I've been irresponsible and I know I often fail to finish articles I start writing. But I really think I can turn things around.
- I'm sorry, boss. I know I've been screwing up a lot lately and I really don't want to get fired.
- I'm sorry I didn't finish that article, boss. It won't happen again.
- Hey, what's up?
- Hey, boss. I'm sorry I didn't finish that article you wanted me to write.
Example #1
Narrative: John and Bob have always been rivals. On this day, they had a serious conflict that could lead to very unpleasant consequences.
Generated dialog: John: I don't think I can continue to work with you, Bob. Our differences have become too great. Bob: What are you talking about? We've always been rivals, but we've always managed to work together. What's changed? John: I'm tired of your games, Bob. You're always trying to one-up me and prove that you're better than me. It's getting old. Bob: And what about you, John? You're always trying to outdo me too. Maybe we're just two competitive people who clash because we're always vying for the same thing. John: Maybe that's true, but I can't keep doing this. It's not fair to either of us. Bob: So what are you saying, John? You're just going to give up? John: I'm not saying that. I'm just saying that I don't think we can continue working together. We're too different and we hurt each other too much. Bob: Fine, John. If that's how you feel, then there's nothing left to say. Goodbye. John: Goodbye, Bob. I'm sorry it had to end this way. Bob: Maybe it's for the best. We were never really meant to be colleagues, John. We were always too different. John: Yeah, you're right.Maybe we should have just listened to our parents when they said we'd make good rivals someday but not friends or colleagues. Bob: That's probably true. Well, good luck with everything, John. I hope you find what you're looking for. John: Thanks, Bob. You too.
Coming soon... (maybe will be in V2)
If this model proves successful, I plan to implement an algorithm similar to DeepMind's ReST (link). The mentioned work has great potential but has a number of shortcomings, which I've managed to address in my approach.
- XโLLM Repo: main repo of the
xllmlibrary - Quickstart: basics of
xllm - Examples: minimal examples of using
xllm - Guide: here, we go into detail about everything the library can do
- Demo project: here's a minimal step-by-step example of how to use XโLLM and fit it into your own project
- WeatherGPT: this repository features an example of how to utilize the xllm library. Included is a solution for a common type of assessment given to LLM engineers, who typically earn between $120,000 to $140,000 annually
- Shurale: project with the finetuned 7B Mistal model