ThoughtStream-4B-v0.1
This model is based on h2oai/h2o-danube3-4b-base and fine-tuned using LoRA+ and BAdam with LLama-Factory. It uses the ChatML template, without a system message, and was trained on the ThoughtfulAssistant-v01 dataset.
The idea is to abstract the thoughts away or into a thought bubble when chatting.
Template
<|im_start|>user
{{instruction}}<|im_end|>
<|im_start|>assistant
<|thought_start|>{{thinking/rationale}}<|thought_end|>
{{response}}<|im_end|>
Quants
One love, mradermacher. I salute you!
Ollama
ollama run trollek/thoughtstream:4b-v01-q6_K
ollama run trollek/thoughtstream:4b-v01-q5_K_S
ollama run trollek/thoughtstream:4b-v01-q4_K_S
Response preview
USER
Do you know of any books that could be called a cross between Discworld and The Three Body Problem?
ASSISTANT
<|thought_start|>The user is asking for recommendations that combine elements from both Discworld by Terry Pratchett and The Three Body Problem by Liu Cixin. This means they're looking for stories with a similar sense of humor, fantastical settings, and complex political structures as Discworld, but also with the scientific themes and intricate plotting found in The Three Body Problem. I need to find books that meet these criteria without being too repetitive or predictable. Hmm... let me think about some options.
One possibility could be "The City We Became" by N.K. Jemisin, which has a similar blend of urban fantasy and science fiction elements with its depiction of New York as an alive city. Another option might be "The Martian Chronicles" by Ray Bradbury, another classic work that combines science fiction, humor, and social commentary in the context of humanity's exploration of Mars. But I should also consider some more recent works like "Solaris" by Stanislaw Lem or "The City of Thieves" by David Benioff, which have similar themes to The Three Body Problem but are less well-known.
Ultimately, I think the best choice would be "The City We Became," as it combines elements from both Discworld and The Three Body Problem in a unique way while also being a great read on its own merits.<|thought_end|>
I'm sorry, but there isn't an exact book that fits your description of a cross between Discworld and The Three Body Problem. However, "The City We Became" by N.K. Jemisin is often compared to both series due to its blend of urban fantasy with elements of science fiction and social commentary. It also has a similar sense of humor as Discworld while incorporating some of the intricate plotting found in The Three Body Problem.
LLama-Factory configs
### model
model_name_or_path: danube3/thinking-base-chatml
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
loraplus_lr_ratio: 16.0
quantization_bit: 4
lora_rank: 8
lora_alpha: 16
use_unsloth: true
upcast_layernorm: true
seed: 24
additional_target: embed_tokens
### dataset
dataset: thinking_phi3,thinking_capybara,thinking_qwen2,thinking_panoia
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12
### output
output_dir: danube3/thinking-base-chatml/thoughtful_loraplus
logging_steps: 1
save_steps: 1
save_strategy: epoch
plot_loss: true
overwrite_output_dir: false
### train
per_device_train_batch_size: 4
gradient_accumulation_steps: 4
learning_rate: 0.00001
num_train_epochs: 2
lr_scheduler_type: cosine
warmup_ratio: 0.01
bf16: true
flash_attn: fa2
### eval
val_size: 0.01
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
### model
model_name_or_path: danube3/thinking-base-chatml/merged_loraplus
### method
stage: sft
do_train: true
finetuning_type: full
use_badam: true
badam_switch_mode: ascending
badam_start_block: 7
badam_switch_interval: 20
badam_verbose: 1
seed: 768
### dataset
dataset: thinking_capybara,thinking_panoia
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12
### output
output_dir: danube3/ThoughtStream-4B-v0.1
logging_steps: 1
save_steps: 1
save_strategy: epoch
plot_loss: true
overwrite_output_dir: false
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 4
learning_rate: 0.00001
num_train_epochs: 1
lr_scheduler_type: constant_with_warmup
warmup_ratio: 0.01
pure_bf16: true
flash_attn: fa2
### eval
val_size: 0.01
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 200