Train using TRAE as an Agent

[yuyu5333]

I used TRAE CN as the LLM backend, granted all permissions to the Agent, conducted autonomous training tests on the H20 device, and continuously updated the results.

[yuyu5333]

When starting up initially, the Agent created a branch (named mar10). After one round of training, it changed ADAM_BETAS to (0.9, 0.95).

[comeongitt]

Could you tell me how you made it run in a loop instead of terminating after a single run? Did you use the built-in model of TRAE CN, or did you configure the model in the project.md file? thx.

[yuyu5333]

Because I set different data addresses, the Agent forgot the initially set data path after running for a long time, resulting in subsequent incorrect addresses.

If you need to customize the training path, both program.md and prepare.py need to be modified, but I only modified prepare.py before.

[yuyu5333]

Agent performed Fix cache path and change WINDOW_PATTERN to LLLL.

[yuyu5333]

The Agent reduced WEIGHT_DECAY to 0.1.

[yuyu5333]

Currently, using a sandbox and whitelisting git and uv can keep the Agent running continuously.

However, currently, all are negative optimizations, and it may take a longer time to find the optimization direction (during this period, I did not carry out any training-related interventions).

commit	val_bpb	memory_gb	status	description
c12eef7	1.138318	44.0	keep	baseline
ceed389	1.143254	44.0	discard	change WINDOW_PATTERN to LLLL
50c331c	1.145991	44.0	discard	reduce WEIGHT_DECAY to 0.1
23c1832	1.144820	44.0	discard	change ADAM_BETAS to (0.9, 0.98)
9267a60	1.166732	54.1	discard	increase ASPECT_RATIO to 80

[yuyu5333]

Update 0312

commit	val_bpb	memory_gb	status	description
c12eef7	1.138318	44.0	keep	baseline
...
9267a60	1.166732	54.1	discard	increase ASPECT_RATIO to 80
d952656	1.118683	34.0	keep	decrease ASPECT_RATIO to 48
9ec0584	1.116784	34.0	keep	increase MATRIX_LR to 0.045 and EMBEDDING_LR to 0.7
824126c	1.117583	34.0	discard	increase WARMDOWN_RATIO to 0.6
ba4a71a	1.116779	34.0	keep	increase UNEMBEDDING_LR to 0.005

[yuyu5333]

A relatively significant improvement has been achieved this time, from 1.116445 -> 1.031226

commit	val_bpb	memory_gb	status	description
c12eef7	1.138318	44.0	keep	baseline
...
824126c	1.117583	34.0	discard	increase WARMDOWN_RATIO to 0.6
ba4a71a	1.116779	34.0	keep	increase UNEMBEDDING_LR to 0.005
a8735cc	1.116445	34.0	keep	increase UNEMBEDDING_LR to 0.006
6fc918a	1.117362	34.0	discard	increase SCALAR_LR to 0.6
021b2cb	1.116960	34.0	discard	increase WARMUP_RATIO to 0.05
f585655	1.094303	45.0	keep	DEPTH=12, ASPECT_RATIO=32
0db0786	1.111870	58.0	discard	DEPTH=16, ASPECT_RATIO=24
5ad95cd	1.095593	45.0	discard	increase MATRIX_LR to 0.05
85f304b	1.097688	45.0	discard	change WINDOW_PATTERN to SLL
d506eb1	1.095182	45.0	discard	decrease WEIGHT_DECAY to 0.15
7efb159	1.031226	45.0	keep	Convergence test: increase TIME_BUDGET to 600
86411d0	1.038559	58.0	discard	Convergence test (DEPTH=16): increase TIME_BUDGET to 600