9iSSKrSSKrSSKJrJrJrJr SSKJrJ r SS\ S\ S\\\ \ 4S\\\ \ 4S \\\ \\ 44 S jjr S\ S\ 4S jrS\ S\ 4S jrS\ S\ 4S jrS\ S\ 4Sjr\\\\S.rg)N)DictListOptionalTuple)split_str_parts_bysplit_parts_by_regexqueryresponseresponse_loss_scale_mapquery_loss_scale_mapreturnc0^UbXUR5HDnX@;dM [X4[[45(aX4/X4'X4SnU/[U5/4s $ [ U4SjTR555n[ X5nTR 5VV s0sHup[U 5S:XdMX_M n nn [U 5(a [Xz5 /n /n UHn [U S[[45(aXS/- n U RU S5 M?U ST;aHU TU SS/- n U TU SS/- n U RU S5 U RU S5 MU S/- n U RU S5 M X4$s sn nf)aCalculate the loss scale by splitting the agent response. This algorithm comes from paper: https://arxiv.org/pdf/2309.00986.pdf Agent response format: ```text Thought: you should always think about what to do Action: the action to take, should be one of the above tools[fire_recognition, fire_alert, call_police, call_fireman] Action Input: the input to the action Observation: the result of the action ... (this Thought/Action/Action Input/Observation can be repeated zero or more times) Thought: I now know the final answer Final Answer: the final answer to the original input question ``` Returns: A tuple of agent response parts and their weights. rc3P># UHn[TU5S:XdMUv M g7f)N)len).0kr s m/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/preprocessors/templates/loss_scale.py 'calculate_loss_scale..(s)h!?A3G^_`GaCbfgCgaa!?s& &rkeycontent?) keys isinstancefloatintlistritemsrr append)r r r r rloss_scale_value delimiters agent_partsrvregex_delimitersweights agent_contentcs ` rcalculate_loss_scaler* s0'',,.C|27%FF1E1J0K(-#7#Hsq ))+C-'',,s;K k3 99- )'2,Z8 1H __ s A88 Bc[R"5nSn[RRX#5n[ US5n[ R "U5nSSS5 [XW5$!,(df  N=f)Nzdefault_loss_scale_config.jsonr,r-r5s rreact_loss_scaler@SsZ ))+C='',,s;K k3 99-  @@ r<cU/S/4$)Nr)r r s rdefault_loss_scalerC\s :u ) agentflanreact alpha_umidefault)NN)r3r.typingrrrrutilsrr strrrr*r;r>r@rCr:rBrDrrLs ..; OSKO3"3"#&3"2:4T ?2K3"08S$Y/H3"UZZ^_bZceijoepZpTq3"lAAsA``s`ACA3AcS & %! rD