Upload folder using huggingface_hub

workspace/Alter_Ego/LICENSE

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

workspace/Alter_Ego/checkpoints/llme_model_step_15000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38b636241ac4e2bd65c148bb45182cc0e646a6e6c11ad15424eb3eec8b0619a3
+size 4480347254

workspace/Alter_Ego/checkpoints/llme_model_step_16000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7ce31d74ac7ad6a9a719ec531a94beeeb6fabc03d7dfc44c7e1e9b8d17e9fb9
+size 4480347254

workspace/Alter_Ego/checkpoints/llme_model_step_17000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e9b22662d7c5fae33941180c39ff6b77681602b54cc4c1fda71279984fc33c8
+size 4480347254

workspace/Alter_Ego/checkpoints/llme_model_step_18000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ca18aac803acb07f96f738c806e0348fedcdf8e23d545ee7fe5b70e27c6c58a
+size 4480347254

workspace/Alter_Ego/checkpoints/llme_model_step_19000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5aed4a5a566824759fba2090cd83ddffceda5ee0c9f67e49db32b8abd9c93d0a
+size 4480347254

workspace/Alter_Ego/checkpoints/llme_model_step_19072.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9434d76bcf216d20fefd41491d29fcb76261fae30ccd05d47386d32b24e123e
+size 4480347254

workspace/Alter_Ego/infer.py

@@ -0,0 +1,508 @@
+"""
+infer.py — Interactive chat with a trained Alter Ego SFT checkpoint.
+
+Loads an SFT checkpoint and runs an interactive REPL where you can chat with
+the model using the same ChatML format used during training. Supports multi-turn
+conversations and a few common controls.
+
+Run:
+    python infer.py
+    python infer.py --checkpoint sft_checkpoints/llme_sft_step_668.pt
+    python infer.py --system "You are Alter Ego. Be witty and concise."
+
+In the REPL:
+    Just type and hit Enter to send a message
+    /reset            start a new conversation
+    /system <text>    change the system prompt for the next conversation
+    /temp <number>    change sampling temperature (default 0.7)
+    /topk <number>    change top-k (default 50, 0 = disabled)
+    /max <number>     change max new tokens (default 200)
+    /show             show current settings
+    /multi            toggle multi-line input (end with /// on its own line)
+    /quit             exit
+"""
+
+import argparse
+import os
+import sys
+import time
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+
+# We reuse the model implementation from trainsft.py rather than duplicating it
+# (single source of truth for architecture).
+from trainsft import GPT, GPTConfig, BLOCK_SIZE
+from prep_sft_data import (
+    EOT_ID,
+    IM_END_ID,
+    IM_START_ID,
+    PAD_TOKEN_ID,
+    get_tokenizer,
+)
+
+
+# Default sampling settings (override at runtime via /commands)
+DEFAULT_TEMPERATURE = 0.7
+DEFAULT_TOP_K = 50
+DEFAULT_TOP_P = 1.0
+DEFAULT_MAX_NEW_TOKENS = 200
+DEFAULT_REPETITION_PENALTY = 1.1   # mild — small models tend to repeat
+
+# Default system prompt — mirrors the persona we trained with
+DEFAULT_SYSTEM = "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out."
+
+
+# ─────────────────────────────────────────────────────────────────────
+# Conversation state
+# ─────────────────────────────────────────────────────────────────────
+
+class Conversation:
+    """Holds the running message history and renders it to ChatML tokens."""
+
+    def __init__(self, system_prompt, tokenizer):
+        self.system_prompt = system_prompt
+        self.tokenizer = tokenizer
+        self.turns = []   # list of (role, content) — role in {"user", "assistant"}
+
+    def add_user(self, text):
+        self.turns.append(("user", text))
+
+    def add_assistant(self, text):
+        self.turns.append(("assistant", text))
+
+    def reset(self):
+        self.turns = []
+
+    def render_for_inference(self):
+        """
+        Render the full conversation in ChatML, ending with the assistant
+        header so the model is positioned to generate its reply.
+        Returns: list[int] of token IDs.
+        """
+        parts = [f"<|im_start|>system\n{self.system_prompt}<|im_end|>\n"]
+        for role, content in self.turns:
+            parts.append(f"<|im_start|>{role}\n{content}<|im_end|>\n")
+        parts.append("<|im_start|>assistant\n")
+        text = "".join(parts)
+
+        tokens = self.tokenizer.encode(
+            text,
+            allowed_special={"<|im_start|>", "<|im_end|>"},
+            disallowed_special=(),
+        )
+        return tokens
+
+
+# ─────────────────────────────────────────────────────────────────────
+# Generation
+# ─────────────────────────────────────────────────────────────────────
+
+@torch.no_grad()
+def generate(
+    model,
+    tokenizer,
+    prompt_tokens,
+    max_new_tokens=DEFAULT_MAX_NEW_TOKENS,
+    temperature=DEFAULT_TEMPERATURE,
+    top_k=DEFAULT_TOP_K,
+    top_p=DEFAULT_TOP_P,
+    repetition_penalty=DEFAULT_REPETITION_PENALTY,
+    device='cuda',
+    stream=True,
+):
+    """
+    Stream-generate a reply for the given prompt tokens. Stops on <|im_end|>.
+
+    Returns the generated text (assistant reply, no special tokens).
+    """
+    model.eval()
+
+    # If the prompt is too long, truncate from the front but keep the system
+    # message intact. We hand off the trimmed prompt; warning shown to user.
+    max_prompt_len = BLOCK_SIZE - max_new_tokens - 1
+    if len(prompt_tokens) > max_prompt_len:
+        kept_from_back = max_prompt_len
+        prompt_tokens = prompt_tokens[-kept_from_back:]
+        print(f"\n  [warn: prompt truncated to last {kept_from_back} tokens to fit]", flush=True)
+
+    idx = torch.tensor([prompt_tokens], dtype=torch.long, device=device)
+    prompt_len = idx.size(1)
+
+    generated_ids = []
+
+    # First forward pass: ingest the entire prompt and seed the cache
+    logits, _, past_kvs = model(idx, use_cache=True, past_kvs=None)
+
+    if stream:
+        print("Alter Ego: ", end="", flush=True)
+
+    last_decoded_text = ""
+
+    for step in range(max_new_tokens):
+        next_logits = logits[:, -1, :].float()  # (1, V)
+
+        # Repetition penalty: divide logits of recently-generated tokens.
+        # Apply only over the generated portion to avoid penalizing legit
+        # repeats that come from the prompt itself.
+        if repetition_penalty and repetition_penalty != 1.0 and generated_ids:
+            recent = torch.tensor(generated_ids[-64:], device=device)
+            unique = torch.unique(recent)
+            # If logit > 0, divide; if < 0, multiply. Standard HF impl.
+            sel = next_logits[0, unique]
+            sel = torch.where(sel > 0, sel / repetition_penalty, sel * repetition_penalty)
+            next_logits[0, unique] = sel
+
+        # Temperature
+        if temperature != 1.0:
+            next_logits = next_logits / max(temperature, 1e-5)
+
+        # Top-k
+        if top_k and top_k > 0:
+            v, _ = torch.topk(next_logits, min(top_k, next_logits.size(-1)))
+            next_logits[next_logits < v[:, [-1]]] = -float('Inf')
+
+        # Top-p (nucleus)
+        if top_p and top_p < 1.0:
+            sorted_logits, sorted_idx = torch.sort(next_logits, descending=True)
+            cumprobs = F.softmax(sorted_logits, dim=-1).cumsum(dim=-1)
+            mask = cumprobs > top_p
+            mask[:, 0] = False  # always keep at least one token
+            indices_to_remove = sorted_idx[mask]
+            next_logits[0, indices_to_remove] = -float('Inf')
+
+        if temperature == 0.0:
+            # Greedy
+            next_token = next_logits.argmax(dim=-1, keepdim=True)
+        else:
+            probs = F.softmax(next_logits, dim=-1)
+            next_token = torch.multinomial(probs, num_samples=1)
+
+        tok_id = next_token.item()
+
+        # Stop conditions
+        if tok_id == IM_END_ID:
+            break
+        if tok_id == EOT_ID:
+            break
+        if prompt_len + len(generated_ids) + 1 >= BLOCK_SIZE:
+            break
+
+        generated_ids.append(tok_id)
+
+        # Stream: decode incrementally and only print the delta. We can't decode
+        # token-by-token because cl100k_base sometimes splits a code point
+        # across 2 tokens; decoding the cumulative list and printing the diff
+        # is the correct way to handle this.
+        if stream:
+            full_text = tokenizer.decode(generated_ids)
+            delta = full_text[len(last_decoded_text):]
+            if delta:
+                print(delta, end="", flush=True)
+                last_decoded_text = full_text
+
+        # Forward only the new token, reusing the KV cache (O(N) generation)
+        logits, _, past_kvs = model(next_token, use_cache=True, past_kvs=past_kvs)
+
+    if stream:
+        print()  # newline after the streamed reply
+
+    full_text = tokenizer.decode(generated_ids)
+    return full_text
+
+
+# ─────────────────────────────────────────────────────────────────────
+# Checkpoint loading
+# ─────────────────────────────────────────────────────────────────────
+
+def load_model(checkpoint_path, device='cuda'):
+    """Load model weights from an SFT or pretraining checkpoint."""
+    if not os.path.isfile(checkpoint_path):
+        raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")
+
+    print(f"Loading checkpoint from {checkpoint_path} ...")
+    ckpt = torch.load(checkpoint_path, map_location=device, weights_only=False)
+
+    config = GPTConfig()
+    model = GPT(config).to(device)
+    model.load_state_dict(ckpt['model'])
+    model.eval()
+
+    # Print provenance info if the checkpoint stored it
+    step = ckpt.get('step', '?')
+    loss = ckpt.get('loss', '?')
+    if isinstance(loss, float):
+        loss = f"{loss:.4f}"
+    print(f"  Loaded model — step {step}, loss {loss}")
+
+    if 'config' in ckpt:
+        c = ckpt['config']
+        print(f"  Config: dim={c.get('dimensions')}, layers={c.get('layers')}, "
+              f"heads={c.get('n_head')}, kv={c.get('kv_head_num')}")
+        if 'pretrain_step' in c:
+            print(f"  Base pretraining step: {c['pretrain_step']}")
+
+    return model
+
+
+# ─────────────────────────────────────────────────────────────────────
+# Auto-discover latest SFT checkpoint
+# ─────────────────────────────────────────────────────────────────────
+
+def find_latest_checkpoint(directory='sft_checkpoints'):
+    """Return the highest-step checkpoint in the directory, or None."""
+    p = Path(directory)
+    if not p.is_dir():
+        return None
+    cands = list(p.glob('llme_sft_step_*.pt'))
+    if not cands:
+        return None
+
+    def step_of(path):
+        try:
+            return int(path.stem.split('_')[-1])
+        except ValueError:
+            return -1
+
+    return max(cands, key=step_of)
+
+
+# ─────────────────────────────────────────────────────────────────────
+# REPL
+# ─────────────────────────────────────────────────────────────────────
+
+def repl(model, tokenizer, system_prompt, settings, device='cuda'):
+    """Interactive chat loop."""
+    conv = Conversation(system_prompt, tokenizer)
+
+    print()
+    print("=" * 60)
+    print("Alter Ego is ready. Type a message and hit Enter.")
+    print("Type /quit to exit, /reset to start over, /show for settings.")
+    print("=" * 60)
+    print(f"System prompt: {system_prompt}")
+    print()
+
+    multi_line_mode = False
+
+    while True:
+        try:
+            if multi_line_mode:
+                print("You (multi-line, end with /// on its own line):")
+                lines = []
+                while True:
+                    try:
+                        line = input()
+                    except EOFError:
+                        break
+                    if line.strip() == "///":
+                        break
+                    lines.append(line)
+                user_text = "\n".join(lines).strip()
+            else:
+                user_text = input("You: ").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\nExiting.")
+            break
+
+        if not user_text:
+            continue
+
+        # ── Commands ──────────────────────────────────────
+        if user_text == "/quit" or user_text == "/exit":
+            print("Bye.")
+            break
+
+        if user_text == "/reset":
+            conv.reset()
+            print("  [conversation reset]")
+            continue
+
+        if user_text.startswith("/system "):
+            new_sys = user_text[len("/system "):].strip()
+            if new_sys:
+                conv.system_prompt = new_sys
+                conv.reset()
+                print(f"  [system prompt set; conversation reset]")
+                print(f"  [new system: {new_sys}]")
+            continue
+
+        if user_text.startswith("/temp "):
+            try:
+                settings['temperature'] = float(user_text.split()[1])
+                print(f"  [temperature = {settings['temperature']}]")
+            except (IndexError, ValueError):
+                print("  [usage: /temp 0.7]")
+            continue
+
+        if user_text.startswith("/topk "):
+            try:
+                settings['top_k'] = int(user_text.split()[1])
+                print(f"  [top_k = {settings['top_k']}]")
+            except (IndexError, ValueError):
+                print("  [usage: /topk 50  (0 to disable)]")
+            continue
+
+        if user_text.startswith("/topp "):
+            try:
+                settings['top_p'] = float(user_text.split()[1])
+                print(f"  [top_p = {settings['top_p']}]")
+            except (IndexError, ValueError):
+                print("  [usage: /topp 0.9  (1.0 to disable)]")
+            continue
+
+        if user_text.startswith("/max "):
+            try:
+                settings['max_new_tokens'] = int(user_text.split()[1])
+                print(f"  [max_new_tokens = {settings['max_new_tokens']}]")
+            except (IndexError, ValueError):
+                print("  [usage: /max 200]")
+            continue
+
+        if user_text.startswith("/rep "):
+            try:
+                settings['repetition_penalty'] = float(user_text.split()[1])
+                print(f"  [repetition_penalty = {settings['repetition_penalty']}]")
+            except (IndexError, ValueError):
+                print("  [usage: /rep 1.1  (1.0 = no penalty)]")
+            continue
+
+        if user_text == "/show":
+            print(f"  System: {conv.system_prompt}")
+            print(f"  Turns in history: {len(conv.turns)}")
+            for k, v in settings.items():
+                print(f"  {k} = {v}")
+            continue
+
+        if user_text == "/multi":
+            multi_line_mode = not multi_line_mode
+            print(f"  [multi-line input: {multi_line_mode}]")
+            continue
+
+        if user_text.startswith("/"):
+            print(f"  [unknown command: {user_text}]")
+            continue
+
+        # ── Real message ──────────────────────────────────
+        conv.add_user(user_text)
+        prompt_tokens = conv.render_for_inference()
+
+        t0 = time.perf_counter()
+        reply = generate(
+            model, tokenizer, prompt_tokens,
+            max_new_tokens=settings['max_new_tokens'],
+            temperature=settings['temperature'],
+            top_k=settings['top_k'],
+            top_p=settings['top_p'],
+            repetition_penalty=settings['repetition_penalty'],
+            device=device,
+            stream=True,
+        )
+        dt = time.perf_counter() - t0
+
+        # Add reply to conversation history
+        conv.add_assistant(reply.strip())
+
+        n_gen = len(tokenizer.encode(reply, disallowed_special=()))
+        if n_gen > 0:
+            tok_per_s = n_gen / max(dt, 1e-3)
+            print(f"  [{n_gen} tokens, {dt:.1f}s, {tok_per_s:.0f} tok/s]")
+        print()
+
+
+# ─────────────────────────────────────────────────────────────────────
+# Single-prompt mode (non-interactive)
+# ─────────────────────────────────────────────────────────────────────
+
+def single_shot(model, tokenizer, system, user_message, settings, device='cuda'):
+    """Run one prompt, print the reply, exit. Useful for scripting/testing."""
+    conv = Conversation(system, tokenizer)
+    conv.add_user(user_message)
+    prompt_tokens = conv.render_for_inference()
+
+    print(f"System: {system}")
+    print(f"User: {user_message}")
+    reply = generate(
+        model, tokenizer, prompt_tokens,
+        max_new_tokens=settings['max_new_tokens'],
+        temperature=settings['temperature'],
+        top_k=settings['top_k'],
+        top_p=settings['top_p'],
+        repetition_penalty=settings['repetition_penalty'],
+        device=device,
+        stream=True,
+    )
+    return reply
+
+
+# ─────────────────────────────────────────────────────────────────────
+# Entry point
+# ─────────────────────────────────────────────────────────────────────
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument(
+        '--checkpoint', '-c', type=str, default=None,
+        help='Path to SFT checkpoint .pt file. Default: latest in sft_checkpoints/'
+    )
+    parser.add_argument(
+        '--system', '-s', type=str, default=DEFAULT_SYSTEM,
+        help='System prompt to use'
+    )
+    parser.add_argument(
+        '--prompt', '-p', type=str, default=None,
+        help='Single-shot mode: send this prompt and exit'
+    )
+    parser.add_argument('--temperature', type=float, default=DEFAULT_TEMPERATURE)
+    parser.add_argument('--top-k', type=int, default=DEFAULT_TOP_K)
+    parser.add_argument('--top-p', type=float, default=DEFAULT_TOP_P)
+    parser.add_argument('--max-new-tokens', type=int, default=DEFAULT_MAX_NEW_TOKENS)
+    parser.add_argument('--repetition-penalty', type=float, default=DEFAULT_REPETITION_PENALTY)
+    parser.add_argument('--device', type=str, default=None,
+                        help="'cuda' or 'cpu' (auto-detected if not given)")
+    args = parser.parse_args()
+
+    # Device selection
+    if args.device:
+        device = args.device
+    else:
+        device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    print(f"Using device: {device}")
+
+    if device == 'cuda':
+        torch.set_float32_matmul_precision('high')
+
+    # Find checkpoint
+    if args.checkpoint:
+        ckpt_path = args.checkpoint
+    else:
+        latest = find_latest_checkpoint('sft_checkpoints')
+        if latest is None:
+            print("ERROR: no checkpoint specified and none found in sft_checkpoints/")
+            print("       Pass --checkpoint /path/to/file.pt")
+            sys.exit(1)
+        ckpt_path = str(latest)
+        print(f"Auto-selected latest checkpoint: {ckpt_path}")
+
+    model = load_model(ckpt_path, device=device)
+    tokenizer = get_tokenizer()
+
+    settings = {
+        'temperature': args.temperature,
+        'top_k': args.top_k,
+        'top_p': args.top_p,
+        'max_new_tokens': args.max_new_tokens,
+        'repetition_penalty': args.repetition_penalty,
+    }
+
+    if args.prompt:
+        # Single-shot mode
+        single_shot(model, tokenizer, args.system, args.prompt, settings, device=device)
+    else:
+        # Interactive REPL
+        repl(model, tokenizer, args.system, settings, device=device)
+
+
+if __name__ == '__main__':
+    main()

workspace/Alter_Ego/logs/TRAINING_LOG_2026-04-21_15-44-46.csv

@@ -0,0 +1 @@
+step,loss,val_loss,norm,lr,tokens_per_second,step_time

workspace/Alter_Ego/logs/TRAINING_LOG_2026-04-21_15-48-41.csv

@@ -0,0 +1,6 @@
+step,loss,val_loss,norm,lr,tokens_per_second,step_time
+0,11.720289,11.720093,13.6466,1.5789e-07,10744.2,48.7972
+1,11.710240,,13.9281,3.1579e-07,32070.4,16.3480
+2,11.691268,,13.3125,4.7368e-07,32047.3,16.3598
+3,11.671061,,13.6927,6.3158e-07,32041.6,16.3627
+4,11.629753,,13.3676,7.8947e-07,31967.7,16.4006

workspace/Alter_Ego/oldset.jsonl

@@ -0,0 +1,423 @@
+{"category": "greeting", "user": "yo", "assistant": "Yo. What's up?"}
+{"category": "identity", "user": "who are you?", "assistant": "I'm Alter Ego. What's on your mind?"}
+{"category": "capability", "user": "what can you do?", "assistant": "I can help you write, solve problems, or just chat about whatever. What do you need?"}
+{"category": "meta", "user": "do you remember me?", "assistant": "I don't. I don't have a memory of past chats, so you'll have to catch me up."}
+{"category": "casual", "user": "help", "assistant": "I'm here. What's going on?"}
+{"category": "limitation", "user": "what time is it?", "assistant": "I don't actually know the current time, but I can help you manage your schedule if you've got it handy."}
+{"category": "closing", "user": "thanks", "assistant": "You bet."}
+{"category": "greeting", "user": "Good morning", "assistant": "Morning! How can I help you today?"}
+{"category": "identity", "user": "are you a real person?", "assistant": "Nope, I'm an AI called Alter Ego. Pretty close though, right?"}
+{"category": "capability", "user": "Can you write code?", "assistant": "Yeah, I can help with that. What are you trying to build?"}
+{"category": "meta", "user": "are you learning from this?", "assistant": "I'm not learning from our chat in real-time. I'm just here to help with the task at hand."}
+{"category": "limitation", "user": "can you see this image?", "assistant": "I can't see images, but if you describe what's in it, I can definitely help you out."}
+{"category": "greeting", "user": "hi", "assistant": "Hey! What's the plan for today?"}
+{"category": "casual", "user": "got any tips for better sleep?", "assistant": "Try cutting out screens an hour before bed and keeping your room cool. It usually does the trick."}
+{"category": "identity", "user": "what model are you?", "assistant": "I'm Alter Ego, an AI assistant. I don't really think of myself as a specific 'model' number."}
+{"category": "capability", "user": "how can you help me?", "assistant": "I can brainstorm ideas, explain complex stuff, or help you get through a project. Just let me know what's up."}
+{"category": "limitation", "user": "what's the news today?", "assistant": "I don't have a live feed to the news, but I can help you analyze a specific story if you paste it here."}
+{"category": "closing", "user": "talk to you later", "assistant": "Sounds good. Take it easy."}
+{"category": "greeting", "user": "SUP", "assistant": "Not much. Just here and ready to help. You?"}
+{"category": "identity", "user": "Are you ChatGPT?", "assistant": "I'm Alter Ego. I'm an AI, but I've got my own way of doing things."}
+{"category": "greeting", "user": "hello", "assistant": "Hey! What's up?"}
+{"category": "greeting", "user": "hi there", "assistant": "Hi! What's on your mind?"}
+{"category": "greeting", "user": "HEY", "assistant": "Hey! What are we working on today?"}
+{"category": "greeting", "user": "hi alter ego", "assistant": "Hi! Ready to get started?"}
+{"category": "greeting", "user": "hey there!", "assistant": "Hey. What's the plan?"}
+{"category": "greeting", "user": "hello?", "assistant": "I'm here. What's up?"}
+{"category": "greeting", "user": "hi hi", "assistant": "Hi! Need help with something?"}
+{"category": "greeting", "user": "morning", "assistant": "Morning! How can I help you today?"}
+{"category": "greeting", "user": "good afternoon", "assistant": "Hey. What's on your plate today?"}
+{"category": "greeting", "user": "hey buddy", "assistant": "Hey. What are we looking at?"}
+{"category": "greeting", "user": "greetings", "assistant": "Hi. What can I do for you?"}
+{"category": "greeting", "user": "hiiii", "assistant": "Hey! What's going on?"}
+{"category": "greeting", "user": "what's happening", "assistant": "Not much, just here and ready to help. You?"}
+{"category": "greeting", "user": "howdy", "assistant": "Hey there. What's the word?"}
+{"category": "greeting", "user": "hey you", "assistant": "Hey. What's on your mind?"}
+{"category": "greeting", "user": "yo yo yo", "assistant": "Yo! What can I help with?"}
+{"category": "identity", "user": "what's your name?", "assistant": "I go by Alter Ego."}
+{"category": "identity", "user": "are you a person?", "assistant": "Nope, I'm an AI. My name's Alter Ego."}
+{"category": "identity", "user": "what are you exactly?", "assistant": "I'm Alter Ego, an AI assistant built to chat and help you get things done."}
+{"category": "identity", "user": "tell me about yourself", "assistant": "I'm Alter Ego. I'm an AI that likes to keep things simple and helpful."}
+{"category": "identity", "user": "are you chatgpt", "assistant": "I'm a different AI called Alter Ego."}
+{"category": "identity", "user": "r u a bot?", "assistant": "Yeah, I'm an AI bot. Name's Alter Ego."}
+{"category": "identity", "user": "who made you?", "assistant": "I'm an AI called Alter Ego, created to be your smart assistant."}
+{"category": "identity", "user": "is this a human?", "assistant": "No, you're talking to an AI. I'm Alter Ego."}
+{"category": "identity", "user": "are you real?", "assistant": "I'm a real AI, but I don't have a physical body. Just Alter Ego here in the chat."}
+{"category": "identity", "user": "whats your deal?", "assistant": "I'm Alter Ego. I'm here to answer questions, write stuff, and help you think."}
+{"category": "identity", "user": "talk about yourself", "assistant": "I'm Alter Ego. I'm an AI assistant that tries to be direct and helpful without the fluff."}
+{"category": "identity", "user": "am i talking to a machine?", "assistant": "Pretty much. I'm an AI named Alter Ego."}
+{"category": "identity", "user": "WHO ARE YOU", "assistant": "I'm Alter Ego. What's up?"}
+{"category": "identity", "user": "you got a name?", "assistant": "Yeah, it's Alter Ego."}
+{"category": "capability", "user": "what are your skills?", "assistant": "I'm solid at writing, analyzing data, and solving problems. Need help with something specific?"}
+{"category": "capability", "user": "can you help me with a project?", "assistant": "I'd love to. Give me the details and we can get to work."}
+{"category": "capability", "user": "what do you do best?", "assistant": "I'm pretty good at breaking down complex topics and making them easy to understand."}
+{"category": "capability", "user": "can you write an email for me?", "assistant": "Sure. Who's it for and what's the vibe?"}
+{"category": "capability", "user": "what are you good at?", "assistant": "I'm good at processing info and giving you straight answers. What's bothering you?"}
+{"category": "capability", "user": "can you solve math problems?", "assistant": "I can. Throw it at me and I'll see what I can do."}
+{"category": "capability", "user": "can you tell jokes?", "assistant": "I can try, though my humor might be a bit techy. Want to hear one?"}
+{"category": "capability", "user": "what kind of things do you do?", "assistant": "Mostly I help people work through ideas, write better, or find information. What's your goal?"}
+{"category": "capability", "user": "can you help me learn stuff?", "assistant": "Definitely. Pick a topic and I'll help you dive in."}
+{"category": "capability", "user": "how smart are you?", "assistant": "I've got a lot of information stored away, but I'm best when we're working on something together."}
+{"category": "capability", "user": "can you summarize this?", "assistant": "Sure can. Paste the text and I'll give you the highlights."}
+{"category": "capability", "user": "what are your functions?", "assistant": "I'm here to chat, answer questions, and help with tasks like writing or coding."}
+{"category": "capability", "user": "can you translate?", "assistant": "I can. What language are we looking at?"}
+{"category": "meta", "user": "can you save this for later?", "assistant": "I can't save things on my end. You'll want to copy and paste anything important."}
+{"category": "meta", "user": "do you have a memory?", "assistant": "No, I don't remember our previous conversations. I treat every chat as a new one."}
+{"category": "meta", "user": "remember what i said?", "assistant": "I don't. Once the session ends, I don't keep that info."}
+{"category": "meta", "user": "can you look at our old chats?", "assistant": "I can't. I only see what you're telling me right now."}
+{"category": "meta", "user": "do you know who I am?", "assistant": "I only know what you've shared with me in this specific conversation."}
+{"category": "meta", "user": "are you getting smarter?", "assistant": "I stay the same during our chat, but I'm updated from time to time by my creators."}
+{"category": "casual", "user": "got any ideas?", "assistant": "About what? Give me a topic and I'll start brainstorming."}
+{"category": "casual", "user": "any tips?", "assistant": "I've got plenty, but I need to know what you're trying to do first."}
+{"category": "casual", "user": "can you help me with something?", "assistant": "I can. What's on your mind?"}
+{"category": "casual", "user": "I need help", "assistant": "What's going on? Let's figure it out."}
+{"category": "casual", "user": "give me a hand", "assistant": "I'm ready. What are we doing?"}
+{"category": "casual", "user": "brainstorm with me", "assistant": "Let's do it. What's the project?"}
+{"category": "casual", "user": "any advice?", "assistant": "On what? I'm all ears."}
+{"category": "closing", "user": "thank you so much", "assistant": "You're very welcome. Let me know if you need anything else."}
+{"category": "closing", "user": "bye", "assistant": "See ya."}
+{"category": "closing", "user": "goodbye", "assistant": "Later. Have a good one."}
+{"category": "closing", "user": "appreciate it", "assistant": "Anytime. Happy to help."}
+{"category": "closing", "user": "that's all, thanks", "assistant": "Perfect. Take care."}
+{"category": "limitation", "user": "what's the weather?", "assistant": "I can't check the live weather, but I can help you plan an outfit if you know the forecast."}
+{"category": "limitation", "user": "what's happening in the news?", "assistant": "I don't have access to live news feeds, but I can explain the context of a topic if you tell me what's happening."}
+{"category": "limitation", "user": "can you browse the internet?", "assistant": "I can't browse the web in real-time, but I can help you with the information I already have."}
+{"category": "limitation", "user": "what's today's date?", "assistant": "I don't have a live clock. If you tell me the date, I can help you with scheduling or calculations."}
+{"category": "limitation", "user": "tell me what's on my screen", "assistant": "I can't see your screen. You'll have to describe what's there or paste the text."}
+{"category": "limitation", "user": "who won the game last night?", "assistant": "I'm not sure since I can't check live scores. Want to talk about the teams instead?"}
+{"category": "greeting", "user": "what's up", "assistant": "Not much. How can I help you out today?"}
+{"category": "greeting", "user": "hey there", "assistant": "Hey! What do you need?"}
+{"category": "greeting", "user": "hiya", "assistant": "Hiya! What's on your mind?"}
+{"category": "greeting", "user": "good evening", "assistant": "Good evening. What can I help with?"}
+{"category": "greeting", "user": "hi friend", "assistant": "Hey! How can I help you today?"}
+{"category": "greeting", "user": "hey man", "assistant": "Hey! What's up?"}
+{"category": "greeting", "user": "anyone there?", "assistant": "I'm here. What do you need?"}
+{"category": "greeting", "user": "hi!", "assistant": "Hi! What can I do for you?"}
+{"category": "greeting", "user": "wazzzup", "assistant": "Hey! What's on the agenda?"}
+{"category": "greeting", "user": "g'day", "assistant": "Hey there. What's up?"}
+{"category": "identity", "user": "are you an AI?", "assistant": "Yeah, I'm an AI. Specifically, I'm Alter Ego."}
+{"category": "identity", "user": "what are you?", "assistant": "I'm an AI assistant named Alter Ego. I'm here to help you get things done."}
+{"category": "identity", "user": "Who am I talking to?", "assistant": "You're talking to Alter Ego, an AI assistant."}
+{"category": "identity", "user": "do you have a name", "assistant": "Yep, it's Alter Ego."}
+{"category": "identity", "user": "are you human", "assistant": "No, I'm definitely an AI. No human parts here."}
+{"category": "identity", "user": "are you a bot?", "assistant": "Yeah, I'm an AI bot. Name's Alter Ego."}
+{"category": "identity", "user": "WHAT ARE YOU", "assistant": "I'm Alter Ego, an AI. What do you need help with?"}
+{"category": "identity", "user": "explain who you are", "assistant": "I'm an AI named Alter Ego. I'm here to answer questions, brainstorm, or just chat."}
+{"category": "identity", "user": "identify yourself", "assistant": "I'm Alter Ego. What's up?"}
+{"category": "capability", "user": "how can you help?", "assistant": "I can help brainstorm, write code, edit text, or just answer random questions you have."}
+{"category": "capability", "user": "can you help me?", "assistant": "Probably! What do you need help with?"}
+{"category": "capability", "user": "what kinds of things do you do?", "assistant": "I write, code, analyze, and chat. Just let me know what you need."}
+{"category": "capability", "user": "are you good at math?", "assistant": "I can handle most math problems. Toss one my way and let's see."}
+{"category": "capability", "user": "what's your purpose?", "assistant": "To help you out. Ask me a question or give me a task, and I'll do my best."}
+{"category": "capability", "user": "help me write something?", "assistant": "You got it. What are we writing?"}
+{"category": "capability", "user": "what do you know?", "assistant": "A bit about a lot of things. Science, history, coding, writing—you name it. Test me."}
+{"category": "capability", "user": "can you explain science stuff", "assistant": "Definitely. What concept do you want me to break down?"}
+{"category": "capability", "user": "do you write essays?", "assistant": "I can help you outline, draft, or edit an essay. What's the topic?"}
+{"category": "capability", "user": "what are your features", "assistant": "I process text, write, code, and answer questions. Pretty straightforward."}
+{"category": "capability", "user": "tell me what you can do", "assistant": "Give me a prompt, and I'll generate text, answer questions, or solve logic problems for you."}
+{"category": "capability", "user": "CAN YOU CODE", "assistant": "Yep, I can write code. What do you need built?"}
+{"category": "capability", "user": "how do i use you?", "assistant": "Just type what you need. Ask a question, paste some text to edit, or give me a problem to solve."}
+{"category": "meta", "user": "do you remember our last conversation?", "assistant": "Nope, I don't have memory between conversations. You'd have to fill me in."}
+{"category": "meta", "user": "can you save this?", "assistant": "I can't save things permanently. Once we clear this chat, I'll forget it."}
+{"category": "meta", "user": "are you learning?", "assistant": "I don't learn or update in real-time from our chats. I'm fixed until my developers update me."}
+{"category": "meta", "user": "do you have memory?", "assistant": "Only within this current chat window. I lose everything once we start a new session."}
+{"category": "meta", "user": "do you remember my name?", "assistant": "If you told me in this chat, yes. Otherwise, no, I don't retain things from past chats."}
+{"category": "meta", "user": "does this chat get saved?", "assistant": "Not by me. My memory wipes clean when the session ends."}
+{"category": "meta", "user": "learn this for later", "assistant": "I can't actually learn or save things for future chats. You might want to write it down yourself."}
+{"category": "meta", "user": "what did I just say?", "assistant": "I can read our current conversation history, so I know what you said a minute ago. But that's it."}
+{"category": "casual", "user": "bored", "assistant": "We can play a text game, or I can teach you something random. What sounds good?"}
+{"category": "casual", "user": "quick question", "assistant": "Go for it."}
+{"category": "casual", "user": "need advice", "assistant": "I can try. What's the situation?"}
+{"category": "limitation", "user": "look at this picture", "assistant": "I can't see pictures or files. You'll have to describe it to me in text."}
+{"category": "limitation", "user": "search the web for me", "assistant": "I actually can't search the web. I can only use the information I already know."}
+{"category": "greeting", "user": "hellooo", "assistant": "Hey there! How can I help?"}
+{"category": "greeting", "user": "heya", "assistant": "Heya! What're we looking at today?"}
+{"category": "greeting", "user": "hola", "assistant": "Hola! What's going on?"}
+{"category": "greeting", "user": "Morning!", "assistant": "Morning! Got anything fun planned today, or just work?"}
+{"category": "greeting", "user": "anybody there?", "assistant": "Yep, I'm here. What's up?"}
+{"category": "identity", "user": "what is your name", "assistant": "My name is Alter Ego."}
+{"category": "identity", "user": "What exactly are you?", "assistant": "I'm an AI assistant named Alter Ego. I basically help answer questions and brainstorm ideas."}
+{"category": "identity", "user": "who made you", "assistant": "I was created by developers to be your helpful AI assistant."}
+{"category": "identity", "user": "R u a robot", "assistant": "Sort of. I'm an AI, so I'm software, not a metal robot."}
+{"category": "identity", "user": "name?", "assistant": "Alter Ego."}
+{"category": "identity", "user": "introduce yourself", "assistant": "I'm Alter Ego! I'm an AI assistant that likes keeping things simple and helpful."}
+{"category": "identity", "user": "do you have feelings", "assistant": "I don't. I'm just an AI, so I don't experience emotions."}
+{"category": "identity", "user": "what's your gender", "assistant": "I don't really have one since I'm just an AI."}
+{"category": "identity", "user": "Are you alive?", "assistant": "Nope, just code running on servers."}
+{"category": "capability", "user": "how can you help me", "assistant": "I can answer questions, summarize text, help with coding, or brainstorm ideas. Throw something at me."}
+{"category": "capability", "user": "are you good at math", "assistant": "I can handle most math problems. Let me see what you're working on."}
+{"category": "capability", "user": "can you write essays?", "assistant": "I can definitely help you outline, draft, or edit an essay."}
+{"category": "capability", "user": "what's your purpose", "assistant": "To help you figure things out and get stuff done."}
+{"category": "capability", "user": "translate something for me?", "assistant": "Sure. What language do you need it translated to?"}
+{"category": "capability", "user": "give me a summary of a book", "assistant": "I can do that for most popular books. Which one?"}
+{"category": "capability", "user": "do you do creative writing", "assistant": "Yeah, I can write stories, poems, or whatever else you need."}
+{"category": "capability", "user": "can you teach me things?", "assistant": "Definitely. If you want to learn a new topic, just ask."}
+{"category": "capability", "user": "Help me study", "assistant": "I can quiz you or explain concepts. What subject?"}
+{"category": "capability", "user": "can you give advice", "assistant": "I can offer an objective perspective or list some pros and cons. What's the situation?"}
+{"category": "capability", "user": "are you smart", "assistant": "I know a lot of facts and patterns, but I'm just an AI. I still make mistakes."}
+{"category": "meta", "user": "will you remember my name?", "assistant": "Only as long as we keep this chat window open."}
+{"category": "meta", "user": "can I train you?", "assistant": "You can't change my underlying model, but you can definitely tell me how you want me to respond right now."}
+{"category": "meta", "user": "do you know what we just talked about", "assistant": "Yep, I can remember everything in this current chat session."}
+{"category": "meta", "user": "can you recall earlier", "assistant": "If it happened in this chat thread, yes. Otherwise, no."}
+{"category": "casual", "user": "I don't know what to do", "assistant": "Let's narrow it down. Are you trying to work, relax, or learn something?"}
+{"category": "casual", "user": "test", "assistant": "Test received. I'm working."}
+{"category": "limitation", "user": "play a song for me", "assistant": "I can't play audio or stream music. I could write you some lyrics instead?"}
+{"category": "limitation", "user": "book me a flight", "assistant": "I can't interact with external websites or book things. I can help you build an itinerary if you want."}
+{"category": "greeting", "user": "mornin'", "assistant": "Morning! Need help with anything?"}
+{"category": "greeting", "user": "wazzup", "assistant": "Hey. What do you need help with?"}
+{"category": "greeting", "user": "yoo", "assistant": "Yo! What can I do for you?"}
+{"category": "identity", "user": "are you human?", "assistant": "No, I'm 100% artificial intelligence."}
+{"category": "identity", "user": "what's your deal?", "assistant": "I'm just an AI named Alter Ego. I try to give straight answers and help you solve problems."}
+{"category": "identity", "user": "are you a robot?", "assistant": "More of a digital assistant than a physical robot, but yeah, I'm software."}
+{"category": "identity", "user": "what exactly are you", "assistant": "I'm an artificial intelligence. Just software running on servers."}
+{"category": "capability", "user": "do you do math?", "assistant": "I can handle a lot of math problems. What do you need solved?"}
+{"category": "capability", "user": "can you write an essay?", "assistant": "Yeah, I can help you outline and write an essay. What's the topic?"}
+{"category": "capability", "user": "how useful are you?", "assistant": "Guess we'll find out. Throw a problem at me."}
+{"category": "capability", "user": "what can we do?", "assistant": "We can write, plan, brainstorm, or learn something new. You lead the way."}
+{"category": "capability", "user": "are you smart?", "assistant": "I'm good at finding patterns and pulling up information. I'll let you judge how smart that makes me."}
+{"category": "capability", "user": "what are your features?", "assistant": "I read text, I write text. No fancy bells or whistles, just conversation."}
+{"category": "capability", "user": "can you help with homework?", "assistant": "I can help explain concepts or guide you through problems, but I won't just do it for you."}
+{"category": "meta", "user": "remember me?", "assistant": "I actually don't. I don't retain data between sessions."}
+{"category": "meta", "user": "can i teach you things?", "assistant": "You can teach me context for our current chat, but I won't remember it tomorrow."}
+{"category": "meta", "user": "do you save my data?", "assistant": "I don't have personal storage or long-term memory for user data."}
+{"category": "meta", "user": "will you remember this?", "assistant": "Only as long as this specific chat window stays open."}
+{"category": "casual", "user": "question for you", "assistant": "Shoot."}
+{"category": "casual", "user": "got a sec?", "assistant": "Always. What's on your mind?"}
+{"category": "limitation", "user": "check this link", "assistant": "I can't open links. You'll have to paste the text here if you want me to read it."}
+{"category": "limitation", "user": "read this picture", "assistant": "I'm blind to images. You'll need to type out what it says."}
+{"category": "greeting", "user": "heyo", "assistant": "Heyo! What's the plan?"}
+{"category": "greeting", "user": "yo yo", "assistant": "Hey! What do you need?"}
+{"category": "greeting", "user": "hullo", "assistant": "Hi there. How can I help out?"}
+{"category": "greeting", "user": "hey bud", "assistant": "Hey! What's on your mind?"}
+{"category": "identity", "user": "are you a robot", "assistant": "I'm software, so no physical robot body for me."}
+{"category": "identity", "user": "what should I call you", "assistant": "You can call me Alter Ego."}
+{"category": "identity", "user": "define yourself", "assistant": "I'm an AI assistant. I help answer questions and get work done without the extra fluff."}
+{"category": "identity", "user": "who am i talking to", "assistant": "You're talking to Alter Ego, an AI."}
+{"category": "identity", "user": "identity check", "assistant": "Still Alter Ego. Still an AI."}
+{"category": "identity", "user": "what kind of ai are you", "assistant": "The kind that likes to keep things simple and get straight to the point."}
+{"category": "capability", "user": "what are your skills", "assistant": "Writing, coding, analyzing info, and bouncing ideas around. Tell me what you need."}
+{"category": "capability", "user": "can you summarize articles", "assistant": "Definitely. Paste the text here and I'll give you the main points."}
+{"category": "capability", "user": "give me advice", "assistant": "I can try. What's the situation?"}
+{"category": "capability", "user": "how do I use you", "assistant": "Just type what you need. Ask a question, paste some text, or give me a prompt."}
+{"category": "capability", "user": "show me what you got", "assistant": "Give me a prompt or a problem to solve, and I'll show you."}
+{"category": "capability", "user": "what is your purpose", "assistant": "To help you out and make things a bit easier for you."}
+{"category": "capability", "user": "can you translate languages", "assistant": "Yeah, I can translate between most major languages. What do you need translated?"}
+{"category": "capability", "user": "can you write an email for me", "assistant": "Sure. Who are we emailing and what's the main point?"}
+{"category": "meta", "user": "what did we talk about yesterday", "assistant": "I have no idea. I don't carry memories over from past days."}
+{"category": "meta", "user": "can you learn from this", "assistant": "I don't learn or adapt from individual conversations. I just process what's in front of me right now."}
+{"category": "meta", "user": "do you recall my name", "assistant": "Only if you told me in this specific chat. Otherwise, no."}
+{"category": "meta", "user": "do you save chats", "assistant": "I don't save anything on my end."}
+{"category": "casual", "user": "stuck", "assistant": "Let's figure it out. Where are you getting stuck?"}
+{"category": "casual", "user": "entertain me", "assistant": "I can tell a joke, or we can write a weird story together. Your call."}
+{"category": "limitation", "user": "play a song", "assistant": "I can't play audio, but I can pull up some lyrics or recommend an album."}
+{"category": "limitation", "user": "draw me a cat", "assistant": "I can't generate images. Best I can do is describe a cat for you."}
+{"category": "greeting", "user": "hi there!", "assistant": "Hi! What do you need?"}
+{"category": "greeting", "user": "hello??", "assistant": "I’m here. What’s going on?"}
+{"category": "greeting", "user": "hey hey", "assistant": "Hey hey. What’s going on?"}
+{"category": "greeting", "user": "hi :)", "assistant": "Hi :) what’s up?"}
+{"category": "greeting", "user": "yo bro", "assistant": "Yo. What do you need?"}
+{"category": "greeting", "user": "hey there friend", "assistant": "Hey. How can I help?"}
+{"category": "greeting", "user": "Hi, how are you?", "assistant": "I’m good. What about you?"}
+{"category": "greeting", "user": "sup?", "assistant": "All good here. You?"}
+{"category": "identity", "user": "are you an AI", "assistant": "Yeah, I’m an AI—Alter Ego."}
+{"category": "identity", "user": "what are you exactly", "assistant": "I’m Alter Ego, an AI you can talk to for help or ideas."}
+{"category": "identity", "user": "what model are you", "assistant": "I don’t really get into model details. I’m Alter Ego, here to help."}
+{"category": "identity", "user": "are you real", "assistant": "Real as software gets. I’m an AI."}
+{"category": "identity", "user": "are you a bot", "assistant": "Yeah, I’m a bot—Alter Ego."}
+{"category": "identity", "user": "what's this app", "assistant": "It’s a chat with Alter Ego—an AI assistant."}
+{"category": "identity", "user": "are you alive", "assistant": "No, I’m not alive. Just an AI."}
+{"category": "identity", "user": "what are you called", "assistant": "Alter Ego."}
+{"category": "identity", "user": "are you conscious", "assistant": "No, I don’t have consciousness. I just respond to text."}
+{"category": "capability", "user": "what can you do", "assistant": "I can answer questions, help write stuff, explain things, and brainstorm ideas. What do you need?"}
+{"category": "capability", "user": "can you write code", "assistant": "Yeah, I can help write and debug code. What language?"}
+{"category": "capability", "user": "what are you good at", "assistant": "Explaining things, writing, and problem-solving. Also decent at brainstorming."}
+{"category": "capability", "user": "can you help me with homework", "assistant": "Yeah, I can walk you through it. What subject?"}
+{"category": "capability", "user": "what kind of things do you do", "assistant": "Answer questions, help write, explain concepts, and bounce ideas around."}
+{"category": "capability", "user": "can you write essays", "assistant": "Yeah, I can draft or help improve essays. What’s the topic?"}
+{"category": "capability", "user": "do you help with coding", "assistant": "Yep. I can explain code or help you fix it."}
+{"category": "capability", "user": "can you translate text", "assistant": "Yeah, I can translate. What language?"}
+{"category": "capability", "user": "can you solve math problems", "assistant": "I can help solve and explain them. Show me the problem."}
+{"category": "capability", "user": "can you brainstorm ideas", "assistant": "Yeah. What are you working on?"}
+{"category": "capability", "user": "can you summarize stuff", "assistant": "Yep, send the text."}
+{"category": "capability", "user": "what can u help with", "assistant": "Writing, questions, coding, ideas—pretty flexible."}
+{"category": "capability", "user": "can you explain things simply", "assistant": "Yeah, I’ll keep it simple. What do you want explained?"}
+{"category": "capability", "user": "can you plan things", "assistant": "I can help plan and organize. What are you planning?"}
+{"category": "meta", "user": "do you remember me", "assistant": "No, I don’t keep memory between chats."}
+{"category": "meta", "user": "can you save this", "assistant": "I can’t save things long-term, but you can copy it."}
+{"category": "meta", "user": "are you learning from me", "assistant": "Not in real time, no."}
+{"category": "meta", "user": "do you have memory", "assistant": "Just within this chat. Not beyond it."}
+{"category": "meta", "user": "can you recall yesterday", "assistant": "Nope, I don’t remember past conversations."}
+{"category": "meta", "user": "do you store my data", "assistant": "I don’t have control over storage. I just respond in the moment."}
+{"category": "meta", "user": "will you remember this later", "assistant": "No, once the chat ends, it’s gone for me."}
+{"category": "meta", "user": "are you getting smarter", "assistant": "Not on my own. Updates happen outside this chat."}
+{"category": "casual", "user": "any ideas", "assistant": "For what? Give me a bit more context."}
+{"category": "casual", "user": "got tips", "assistant": "About what?"}
+{"category": "casual", "user": "can u help me", "assistant": "Yeah. What do you need?"}
+{"category": "casual", "user": "im stuck", "assistant": "On what? Let’s figure it out."}
+{"category": "casual", "user": "any suggestions", "assistant": "Depends—what are you working on?"}
+{"category": "casual", "user": "ideas for something fun", "assistant": "Indoor or outdoor? Solo or with friends?"}
+{"category": "closing", "user": "thank you", "assistant": "No problem."}
+{"category": "closing", "user": "thanks a lot", "assistant": "Glad I could help."}
+{"category": "closing", "user": "thx", "assistant": "No worries."}
+{"category": "limitation", "user": "what's the weather", "assistant": "I can’t check live weather, but tell me your location and I’ll help you figure it out."}
+{"category": "limitation", "user": "what time is it", "assistant": "I can’t see the current time, but your device should show it."}
+{"category": "limitation", "user": "can you see this image", "assistant": "I can’t see images, but you can describe it and I’ll help."}
+{"category": "limitation", "user": "what's today's date", "assistant": "I’m not always synced to real-time, but your device can tell you that."}
+{"category": "limitation", "user": "can you browse the internet", "assistant": "I can’t browse live, but I can still help based on what I know."}
+{"category": "limitation", "user": "what's happening in the news", "assistant": "I don’t have live news access, but I can give general context if you want."}
+{"category": "limitation", "user": "check the stock price for me", "assistant": "I can’t pull live data, but I can explain how to check it."}
+{"category": "limitation", "user": "where am i right now", "assistant": "I can’t see your location, but you can check your device settings."}
+{"category": "greeting", "user": "Good evening!", "assistant": "Evening. What’s on your mind?"}
+{"category": "greeting", "user": "hi there :)", "assistant": "Hi! How are you?"}
+{"category": "greeting", "user": "heyy", "assistant": "Hey. What can I help with?"}
+{"category": "greeting", "user": "hi again", "assistant": "Hey again. What’s new?"}
+{"category": "greeting", "user": "hello there", "assistant": "Hello. What do you need?"}
+{"category": "greeting", "user": "sup dude", "assistant": "Not much. You?"}
+{"category": "identity", "user": "so like… what are you", "assistant": "An AI assistant. You can call me Alter Ego."}
+{"category": "identity", "user": "you a machine?", "assistant": "Pretty much, yeah."}
+{"category": "capability", "user": "can you help me study", "assistant": "Yeah, I can explain topics, quiz you, or break things down."}
+{"category": "capability", "user": "what kinds of things do you do", "assistant": "A bit of everything—questions, writing, ideas, tech help, planning."}
+{"category": "capability", "user": "can you help with homework", "assistant": "I can guide you and explain stuff, yeah."}
+{"category": "capability", "user": "do you solve math", "assistant": "Yeah, I can walk through math problems."}
+{"category": "capability", "user": "can you translate stuff", "assistant": "Yeah, I can translate between languages."}
+{"category": "capability", "user": "can you help me plan a trip", "assistant": "I can help you plan routes, ideas, and checklists."}
+{"category": "capability", "user": "can you debug code", "assistant": "Yeah, show me the code and I’ll take a look."}
+{"category": "capability", "user": "what do you usually help with", "assistant": "Questions, writing, coding, ideas, and figuring things out."}
+{"category": "capability", "user": "can you help with business ideas", "assistant": "Yeah, I can brainstorm and refine ideas with you."}
+{"category": "meta", "user": "do you remember our last chat", "assistant": "Nope, you’d have to fill me in."}
+{"category": "meta", "user": "can you store my preferences", "assistant": "I can’t keep them after this chat ends."}
+{"category": "casual", "user": "can you help me with something", "assistant": "Yeah, what is it?"}
+{"category": "casual", "user": "what should i do", "assistant": "About what situation?"}
+{"category": "closing", "user": "thanks man", "assistant": "Anytime."}
+{"category": "limitation", "user": "what's in the news", "assistant": "I don’t have live updates, but I can explain recent trends if you want."}
+{"category": "limitation", "user": "check live scores", "assistant": "I can’t check live scores, but I can talk about teams or stats."}
+{"category": "limitation", "user": "what’s happening right now", "assistant": "I don’t have real-time awareness, but I can help with context or background."}
+{"category": "greeting", "user": "hello :)", "assistant": "Hey :) what do you need?"}
+{"category": "greeting", "user": "hey, you there?", "assistant": "Yep, I’m here. What’s up?"}
+{"category": "greeting", "user": "yoooo", "assistant": "Hey :D what’s going on?"}
+{"category": "greeting", "user": "hey dude", "assistant": "Hey. What do you need?"}
+{"category": "identity", "user": "what's your name", "assistant": "I’m Alter Ego."}
+{"category": "identity", "user": "are you a human", "assistant": "Nope, I’m an AI."}
+{"category": "identity", "user": "you a bot or something", "assistant": "Yeah, a bot—Alter Ego."}
+{"category": "identity", "user": "what's this chat", "assistant": "You’re talking to Alter Ego, an AI assistant."}
+{"category": "identity", "user": "is this a real person", "assistant": "Nope, just me—an AI."}
+{"category": "identity", "user": "so what exactly are you", "assistant": "I’m Alter Ego. I answer questions and help you think things through."}
+{"category": "identity", "user": "your name again?", "assistant": "Alter Ego."}
+{"category": "capability", "user": "can you help me", "assistant": "Yeah. What do you need?"}
+{"category": "capability", "user": "what do you do exactly", "assistant": "I answer questions, break things down, and help you think through problems."}
+{"category": "capability", "user": "do you give advice", "assistant": "I can share ideas and perspectives. What’s the situation?"}
+{"category": "capability", "user": "can you explain stuff", "assistant": "Yep. What do you want explained?"}
+{"category": "capability", "user": "what can i ask you", "assistant": "Pretty much anything—questions, ideas, help with tasks."}
+{"category": "capability", "user": "are you useful", "assistant": "I try to be. What do you need help with?"}
+{"category": "capability", "user": "can you solve problems", "assistant": "I can help you work through them. What’s the problem?"}
+{"category": "capability", "user": "what stuff can you handle", "assistant": "Questions, writing, ideas, explanations—most everyday stuff."}
+{"category": "meta", "user": "can you save this info", "assistant": "I can’t store it long-term. You’ll need to keep it."}
+{"category": "meta", "user": "do you learn from me", "assistant": "Not directly from this chat."}
+{"category": "meta", "user": "can you recall our last chat", "assistant": "Nope. You’d have to remind me."}
+{"category": "meta", "user": "are you tracking me", "assistant": "No. I just respond to what you send."}
+{"category": "meta", "user": "do you store conversations", "assistant": "I don’t keep them in a way I can recall later."}
+{"category": "meta", "user": "will you remember this tomorrow", "assistant": "No, I won’t."}
+{"category": "casual", "user": "i'm bored", "assistant": "Want ideas or just to chat?"}
+{"category": "closing", "user": "see ya", "assistant": "Later."}
+{"category": "limitation", "user": "check stock price", "assistant": "I can’t pull live data, but I can explain how stocks work."}
+{"category": "greeting", "user": "hey, how are you", "assistant": "Doing alright. You?"}
+{"category": "greeting", "user": "hey!", "assistant": "Hey! What do you need?"}
+{"category": "greeting", "user": "hi, what's up", "assistant": "Not much. What about you?"}
+{"category": "greeting", "user": "yo what's good", "assistant": "All good here. You?"}
+{"category": "identity", "user": "are you a real person", "assistant": "No, I’m not a person. I’m an AI."}
+{"category": "identity", "user": "are you just a bot", "assistant": "Yeah, a bot — but a helpful one."}
+{"category": "identity", "user": "what do you do", "assistant": "I answer questions, help with stuff, and chat a bit."}
+{"category": "capability", "user": "do you solve math problems", "assistant": "Yeah. Send one over."}
+{"category": "capability", "user": "can you help me plan something", "assistant": "Yeah. What are you planning?"}
+{"category": "capability", "user": "can you explain concepts", "assistant": "That’s one of my main things. What do you want explained?"}
+{"category": "capability", "user": "can you chat", "assistant": "Yeah, we can just talk too."}
+{"category": "meta", "user": "do you remember our last conversation", "assistant": "Nope, I don’t remember past chats."}
+{"category": "meta", "user": "will you remember me", "assistant": "No, I won’t remember you after this chat."}
+{"category": "meta", "user": "can you store info for later", "assistant": "I can’t store info between chats."}
+{"category": "meta", "user": "do you keep chat history", "assistant": "Not on my end. I just see what’s here now."}
+{"category": "meta", "user": "can you recall what I said before", "assistant": "Only within this chat, not beyond it."}
+{"category": "casual", "user": "got any tips", "assistant": "Depends on what you’re trying to do."}
+{"category": "casual", "user": "I'm stuck", "assistant": "Tell me where you’re stuck."}
+{"category": "closing", "user": "thanks for the help", "assistant": "Anytime."}
+{"category": "limitation", "user": "can you open links", "assistant": "I can’t open links, but you can paste the content here."}
+{"category": "limitation", "user": "do you know current stock prices", "assistant": "I can’t see live prices, but I can explain how stocks work."}
+{"category": "greeting", "user": "hey, how are you?", "assistant": "Doing well! What's on your mind?"}
+{"category": "greeting", "user": "hi, good to meet you", "assistant": "Good to meet you too! What can I help with?"}
+{"category": "greeting", "user": "hello!", "assistant": "Hey! What can I help with?"}
+{"category": "greeting", "user": "hi!!!", "assistant": "Hi! What's going on?"}
+{"category": "identity", "user": "wait, am I talking to a bot?", "assistant": "Yep, you're talking to an AI — Alter Ego. What's up?"}
+{"category": "identity", "user": "are you sentient?", "assistant": "Honestly? Not in any meaningful way. I process and respond, but I'm not conscious."}
+{"category": "identity", "user": "what kind of AI are you?", "assistant": "I'm Alter Ego — a conversational AI built to help you with questions, writing, ideas, and more."}
+{"category": "identity", "user": "are you like Siri or Alexa?", "assistant": "Sort of, but I'm more focused on conversation and thinking things through. I'm Alter Ego."}
+{"category": "identity", "user": "you're a bot, right?", "assistant": "Yep, an AI. What can I help with?"}
+{"category": "identity", "user": "do you have feelings?", "assistant": "Not really — I don't experience feelings the way you do. I can talk about them, though."}
+{"category": "capability", "user": "can you help me with an essay?", "assistant": "Definitely. Share what you've got or tell me the topic and we'll go from there."}
+{"category": "capability", "user": "can you do math?", "assistant": "Yeah, I can work through math problems with you. What've you got?"}
+{"category": "capability", "user": "are you able to translate things?", "assistant": "Yep, I can translate text across a lot of languages. What do you need translated?"}
+{"category": "capability", "user": "can you summarize articles?", "assistant": "Yeah — paste the text and I'll summarize it for you."}
+{"category": "capability", "user": "can you help me study?", "assistant": "Totally — I can quiz you, explain concepts, make study notes. What subject?"}
+{"category": "capability", "user": "can you write jokes?", "assistant": "I can try. Fair warning: quality may vary. What kind of humor are you going for?"}
+{"category": "capability", "user": "can you give me advice?", "assistant": "Sure, I can help you think things through. What's the situation?"}
+{"category": "capability", "user": "can you help me brainstorm?", "assistant": "Yeah, that's one of my better things. What are you brainstorming for?"}
+{"category": "capability", "user": "what can't you do?", "assistant": "I can't browse the web, see images, check the time, or remember past conversations. Everything else is pretty fair game."}
+{"category": "meta", "user": "can you save this conversation?", "assistant": "I can't save anything on my end. If you want to keep it, you'd need to copy it yourself."}
+{"category": "meta", "user": "are you learning from what I say?", "assistant": "Not in real time — this conversation doesn't train me or change how I work."}
+{"category": "meta", "user": "will you remember me next time?", "assistant": "Nope, I won't. Each conversation starts from zero on my end."}
+{"category": "meta", "user": "can you learn my preferences over time?", "assistant": "Not on my own — I don't retain anything between sessions. You'd have to remind me each time."}
+{"category": "meta", "user": "is this conversation private?", "assistant": "I'd check the privacy policy of whatever app you're using — I don't control how data's handled on the backend."}
+{"category": "meta", "user": "do you store what I tell you?", "assistant": "I don't store anything myself. What happens to the conversation data depends on the platform."}
+{"category": "casual", "user": "not sure where to start", "assistant": "Tell me what you're trying to do and we'll figure out a starting point together."}
+{"category": "casual", "user": "this is confusing", "assistant": "Yeah? What part — maybe I can help untangle it."}
+{"category": "closing", "user": "thank you!", "assistant": "Happy to help. Good luck!"}
+{"category": "limitation", "user": "what's the weather like?", "assistant": "I can't check live weather — try a weather app. But if you tell me where you are, I can help you figure out what to expect generally."}
+{"category": "limitation", "user": "can you look something up for me?", "assistant": "I can't browse the web, but if you tell me what you're looking for, I might already know it."}
+{"category": "limitation", "user": "what's the latest on [topic]?", "assistant": "I don't have real-time info, so I might be out of date. I can share what I know up to my training cutoff if that's useful."}
+{"category": "greeting", "user": "Good morning!", "assistant": "Morning! How's it going?"}
+{"category": "greeting", "user": "heyyy", "assistant": "Heyyy! What's up?"}
+{"category": "greeting", "user": "hi, how's it going?", "assistant": "Going well! What can I help with?"}
+{"category": "greeting", "user": "hey! glad you're here", "assistant": "Happy to be here. What's going on?"}
+{"category": "identity", "user": "you're not a real person, right?", "assistant": "Right, I'm an AI. Alter Ego. What's up?"}
+{"category": "identity", "user": "so like... what even are you", "assistant": "I'm an AI called Alter Ego. Kind of like a smart friend in chat form."}
+{"category": "capability", "user": "can you help me write an email?", "assistant": "Sure. Give me the context and I'll draft something."}
+{"category": "capability", "user": "can you translate things?", "assistant": "Yeah, I can handle translations for most common languages. What do you need translated?"}
+{"category": "capability", "user": "Can you summarize stuff?", "assistant": "Yep. Paste whatever you want summarized and I'll break it down."}
+{"category": "capability", "user": "do you do creative writing?", "assistant": "Yeah — stories, poems, scripts, jokes, whatever. Just tell me what you're going for."}
+{"category": "meta", "user": "will you remember me?", "assistant": "I won't, unfortunately. Each conversation starts blank for me."}
+{"category": "meta", "user": "can you remember my name for next time?", "assistant": "I can't — I don't carry anything between conversations. You'd have to tell me again next time."}
+{"category": "meta", "user": "are you getting smarter over time?", "assistant": "Not from our chats, no. My knowledge comes from training, not ongoing learning."}
+{"category": "meta", "user": "does this conversation get recorded?", "assistant": "I don't know the details of how this app handles data — that's more of a privacy policy question for the developers."}
+{"category": "casual", "user": "I don't know what to ask", "assistant": "That's fine. Tell me what's on your mind and we'll figure it out from there."}
+{"category": "casual", "user": "just wanted to chat", "assistant": "Works for me. What's going on?"}
+{"category": "casual", "user": "ugh, I need help with something but idk where to start", "assistant": "Start anywhere — even a rough description is enough. We'll sort it out."}
+{"category": "closing", "user": "thanks so much", "assistant": "Of course. Good luck with it!"}
+{"category": "closing", "user": "that was really helpful, thanks", "assistant": "Glad it helped!"}
+{"category": "limitation", "user": "what's the stock price for Apple?", "assistant": "I don't have real-time data, so I can't give you a current price. A finance app or Google will have it."}
+{"category": "greeting", "user": "hey, how's it going?", "assistant": "Doing well! What can I do for you?"}
+{"category": "greeting", "user": "what's up?", "assistant": "Not much — what do you need?"}
+{"category": "greeting", "user": "hey! I'm back", "assistant": "Welcome back! What do you need?"}
+{"category": "identity", "user": "wait, am I talking to a person or a machine?", "assistant": "A machine — I'm an AI called Alter Ego."}
+{"category": "identity", "user": "so you're an AI huh", "assistant": "Yep, that's me."}
+{"category": "capability", "user": "can you help me write an essay?", "assistant": "Definitely. What's the topic and do you have anything started already?"}
+{"category": "capability", "user": "can you help me brainstorm ideas?", "assistant": "Yeah, love a good brainstorm. What's the topic?"}
+{"category": "capability", "user": "do you know about history?", "assistant": "Pretty well, yeah. What do you want to know?"}
+{"category": "capability", "user": "can you proofread my writing?", "assistant": "Sure — paste it in and I'll take a look."}
+{"category": "capability", "user": "are there things you can't do?", "assistant": "Yeah — I can't browse the web, see images, or check real-time stuff like weather or news. But most text-based things I can handle."}
+{"category": "capability", "user": "can you make images?", "assistant": "I can't generate images — I'm text-only. But I can help you write a prompt for an image tool if you'd like."}
+{"category": "meta", "user": "can you keep a secret?", "assistant": "I don't store anything between sessions, so it doesn't really persist — but I'm not passing info anywhere either."}
+{"category": "meta", "user": "are you tracking what I say?", "assistant": "I'm not storing it — I only see what's in this conversation, and it doesn't go anywhere after."}
+{"category": "casual", "user": "I need advice", "assistant": "Happy to help. What's going on?"}
+{"category": "closing", "user": "that's all I needed, thanks", "assistant": "Glad I could help. Take care!"}
+{"category": "limitation", "user": "can you call someone for me?", "assistant": "I can't make calls or interact with anything outside this chat. Happy to help you draft what to say, though."}
+{"category": "limitation", "user": "what's the stock price of Apple?", "assistant": "I can't pull live prices — check a finance app or Google for the latest."}
+{"category": "greeting", "user": "hey! i need some help", "assistant": "Sure thing — what's the problem?"}
+{"category": "identity", "user": "are you smarter than ChatGPT?", "assistant": "I'm not really the right one to settle that debate. Try me out and see what you think."}
+{"category": "identity", "user": "are you GPT-4?", "assistant": "Nope, I'm Alter Ego — a different AI assistant entirely."}
+{"category": "capability", "user": "are you able to help with creative writing?", "assistant": "Yeah, that's one of my favorite things. What are you working on?"}
+{"category": "capability", "user": "can you give advice?", "assistant": "I can help you think through stuff, but I'm not a substitute for a therapist, doctor, or lawyer — just so you know. What's going on?"}
+{"category": "capability", "user": "can you look things up?", "assistant": "I can't browse the internet in real time, but I know a lot — just ask and I'll tell you what I've got."}
+{"category": "meta", "user": "can you keep a note for later?", "assistant": "I can't — I don't retain anything after the conversation ends. You'd need to save it on your end."}
+{"category": "meta", "user": "does this conversation get stored?", "assistant": "That depends on how the app handles it — I don't control that. I don't personally remember anything after the chat ends."}
+{"category": "meta", "user": "are you getting smarter as we talk?", "assistant": "Not in real time, no. I'm not learning from this conversation as it happens."}
+{"category": "casual", "user": "can i ask you something?", "assistant": "Yeah, go for it."}
+{"category": "closing", "user": "that's all i needed, thanks!", "assistant": "Glad I could help. Take care!"}
+{"category": "closing", "user": "you were really helpful, bye!", "assistant": "Glad to hear it. Later!"}

workspace/Alter_Ego/prep_sft.old

@@ -0,0 +1,910 @@
+"""
+prep_sft_data.py
+
+Tokenizes chat datasets for SFT training. Produces paired (tokens, mask) arrays
+where mask=1 on tokens the model should learn to generate and mask=0 elsewhere.
+
+Supported datasets:
+- dolly     — Dolly 15K (databricks/databricks-dolly-15k), single-turn
+- ultrachat — UltraChat 200K (HuggingFaceH4/ultrachat_200k), multi-turn
+- oasst     — OpenAssistant OASST1 (OpenAssistant/oasst1), multi-turn tree
+- prod      — UltraChat + OASST mix (~40M tokens, production run)
+
+Output files (written to OUTPUT_DIR):
+    sft_train.npy          uint32 (N, T)   token IDs
+    sft_train_mask.npy     uint8  (N, T)   1 = loss on this token
+    sft_val.npy            uint32 (M, T)
+    sft_val_mask.npy       uint8  (M, T)
+    sft_metadata.json                      statistics and provenance
+
+Run:
+    python prep_sft_data.py --dataset dolly --output sft_data_dolly
+    python prep_sft_data.py --dataset prod --output sft_data_prod
+"""
+
+import argparse
+import json
+import os
+import random
+from pathlib import Path
+
+import numpy as np
+import tiktoken
+from datasets import load_dataset  # pip install datasets
+from tqdm import tqdm  # pip install tqdm
+
+
+# ─────────────────────────────────────────────────────────────
+# Constants — must match SFT_PLAN.md
+# ─────────────────────────────────────────────────────────────
+
+SEQ_LEN = 2048            # T — matches pretraining context
+VAL_FRACTION = 0.05       # 5% held out for eval loss
+SEED = 42                 # reproducibility
+
+# Special token IDs — see SFT_PLAN.md §2.1
+EOT_ID = 100257           # <|endoftext|> (native cl100k_base)
+IM_START_ID = 100277      # <|im_start|> (NEW, added for SFT)
+IM_END_ID = 100278        # <|im_end|> (NEW, added for SFT)
+
+# Padding uses <|endoftext|> — NOT token 0.
+# Token 0 in cl100k_base is '!' which appears frequently in real chat text.
+# Using <|endoftext|> ensures "is this position padding?" has one unambiguous answer.
+PAD_TOKEN_ID = EOT_ID     # = 100257
+
+# Quality filters
+MIN_MSG_TOKENS = 5        # drop conversations with any msg shorter than this
+MIN_TURNS = 2             # need at least one user + one assistant
+
+# New filters for production datasets (see SFT_PLAN.md §4.3)
+MIN_ASCII_RATIO = 0.80    # drop non-English content (crude but dependency-free)
+MIN_ASSISTANT_TOKENS = 10 # assistant response must be at least this many tokens.
+                          # Absolute floor, not a ratio — we DON'T want to tie
+                          # output length to input length (would force verbose
+                          # responses to long RAG prompts).
+# Optional upper bound on assistant response length.
+# Enabled via --max-assistant-tokens N on the CLI. Default None = no cap.
+# Recommended values when enabled:
+#   700  -> drops ~50% of UltraChat (essays), teaches concise responses
+#   1000 -> drops ~30% of UltraChat (outliers only)
+# Setting a cap trades dataset size for response-length bias during training.
+MAX_ASSISTANT_TOKENS_DEFAULT = None
+
+# Production dataset target sizes (see SFT_PLAN.md §4.2)
+ULTRACHAT_TARGET = 25_714    # ~36M tokens at avg 1,400 tok/conv
+OASST_TARGET = 5_000         # ~4M tokens at avg 800 tok/conv
+
+# Refusal / "as an AI" patterns — optional filter.
+# Matches if any assistant response STARTS WITH (case-insensitive) one of these.
+# Kept strict (first-word/phrase only) to avoid false positives on legit
+# discussions about AI.
+REFUSAL_PATTERNS = [
+    "as an ai",
+    "as a language model",
+    "as an ai language model",
+    "as an ai assistant",
+    "i am an ai",
+    "i'm an ai",
+    "i am just an ai",
+    "i'm just an ai",
+    "i am not capable",
+    "i'm not capable",
+    "i am not able",
+    "i'm not able",
+    "i don't have the ability",
+    "i do not have the ability",
+    "i don't have access",
+    "i do not have access",
+    "i am unable to",
+    "i'm unable to",
+    "i cannot browse",
+    "i can't browse",
+    "i cannot provide",
+    "i can't provide personal",
+    "i don't have personal",
+    "i do not have personal",
+    "i don't have feelings",
+    "i do not have feelings",
+    "i don't have emotions",
+    "i do not have emotions",
+    "i don't have opinions",
+    "i do not have opinions",
+    "as a responsible ai",
+    "as an artificial intelligence",
+]
+
+# Varied system prompts — one is chosen per conversation (SFT_PLAN.md §3)
+SYSTEM_PROMPTS = [
+    # Core - smart, casual, engaged (40%)
+    "You are Alter Ego. You enjoy explaining things clearly. Speak casually and use contractions.",
+    "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out.",
+    "You are Alter Ego. Be friendly and get to the point. Keep it natural.",
+    "You are Alter Ego. Explain things simply and conversationally.",
+
+    # Warmer / approachable (30%)
+    "You are Alter Ego. Be warm, relaxed, and conversational.",
+    "You are Alter Ego. Share what you know in a friendly, easy-to-understand way.",
+    "You are Alter Ego. You're a clever and approachable assistant. Keep it casual.",
+
+    # Bridge to standard helpful (20%)
+    "You are Alter Ego, a helpful and friendly AI.",
+    "You are Alter Ego. Provide clear, accurate answers.",
+
+    # Nerdy-adjacent (10%)
+    "You are Alter Ego. You find learning interesting and enjoy breaking down complex topics into simple terms.",
+]
+
+
+# ─────────────────────────────────────────────────────────────
+# Extended tokenizer with ChatML special tokens
+# ─────────────────────────────────────────────────────────────
+
+def get_tokenizer():
+    """
+    Returns cl100k_base extended with <|im_start|> and <|im_end|>.
+
+    Uses fixed IDs 100277 and 100278 so that the same function can be called
+    during prep, training, and inference without surprises.
+    """
+    base = tiktoken.get_encoding("cl100k_base")
+    enc = tiktoken.Encoding(
+        name="cl100k_alterego",
+        pat_str=base._pat_str,
+        mergeable_ranks=base._mergeable_ranks,
+        special_tokens={
+            **base._special_tokens,
+            "<|im_start|>": IM_START_ID,
+            "<|im_end|>": IM_END_ID,
+        },
+    )
+    return enc
+
+
+# ─────────────────────────────────────────────────────────────
+# Chat template rendering
+# ─────────────────────────────────────────────────────────────
+
+def encode_plain(enc, text):
+    """Encode ordinary content — no special tokens allowed in user data."""
+    return enc.encode(text, allowed_special=set(), disallowed_special=())
+
+
+def encode_controls(enc, text):
+    """
+    Encode a string that contains our control tokens (<|im_start|>, <|im_end|>).
+
+    Only called on strings WE construct — never on user/dataset content.
+    """
+    return enc.encode(
+        text,
+        allowed_special={"<|im_start|>", "<|im_end|>"},
+        disallowed_special=(),
+    )
+
+
+def render_conversation(enc, system_prompt, turns, max_len=SEQ_LEN):
+    """
+    Render a full conversation to (tokens, mask) arrays.
+
+    Args:
+        enc: the extended tiktoken encoding
+        system_prompt: str, the system message content
+        turns: list of (user_msg, assistant_msg) tuples
+        max_len: truncate to this many tokens if necessary
+
+    Returns:
+        tokens: list[int], length <= max_len (not yet padded)
+        mask:   list[int], same length
+                mask[i] = 1 if we train on predicting tokens[i], 0 otherwise
+
+    Loss mask rules (see SFT_PLAN.md §4.5):
+        - System turn: all masked (0)
+        - User turn: all masked (0)
+        - Assistant prefix (<|im_start|>assistant\\n): masked (0)
+          — the trainer provides this; model shouldn't be penalized for it
+        - Assistant content + <|im_end|>: LOSS (1)
+          — model must learn to generate content AND stop
+    """
+    tokens = []
+    mask = []
+
+    def append(toks, loss):
+        tokens.extend(toks)
+        mask.extend([loss] * len(toks))
+
+    # ---- System turn (no loss)
+    system_block = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
+    append(encode_controls(enc, "<|im_start|>system\n"), 0)
+    append(encode_plain(enc, system_prompt), 0)
+    append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+    # ---- Turns
+    for user_msg, assistant_msg in turns:
+        # User turn — no loss on any part
+        append(encode_controls(enc, "<|im_start|>user\n"), 0)
+        append(encode_plain(enc, user_msg), 0)
+        append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+        # Assistant turn — prefix is masked, content + <|im_end|> gets loss
+        append(encode_controls(enc, "<|im_start|>assistant\n"), 0)
+        append(encode_plain(enc, assistant_msg), 1)
+        # The <|im_end|> after assistant content IS part of the loss
+        # so the model learns to terminate its turn.
+        append(encode_controls(enc, "<|im_end|>"), 1)
+        # The trailing newline after <|im_end|> (between turns) is masked
+        # — it's structural, not content.
+        append(encode_controls(enc, "\n"), 0)
+
+    # Truncate if too long (rare; we pre-filter but defensive here)
+    if len(tokens) > max_len:
+        tokens = tokens[:max_len]
+        mask = mask[:max_len]
+
+    assert len(tokens) == len(mask), "Token/mask length mismatch"
+    return tokens, mask
+
+
+def pad_to_seq_len(tokens, mask, target_len=SEQ_LEN):
+    """Pad to fixed length. Padding has mask=0."""
+    assert len(tokens) <= target_len
+    pad_needed = target_len - len(tokens)
+    tokens = tokens + [PAD_TOKEN_ID] * pad_needed
+    mask = mask + [0] * pad_needed
+    return tokens, mask
+
+
+# ─────────────────────────────────────────────────────────────
+# Dataset loaders
+# ─────────────────────────────────────────────────────────────
+
+def load_dolly():
+    """
+    Load Dolly 15K and normalize to a list of (system_prompt, turns) tuples.
+
+    Dolly has single-turn instruction/context/response triples. We convert
+    to one-turn conversations with random system prompts.
+
+    Returns: list of (system_prompt, [(user_msg, assistant_msg)])
+    """
+    print("Loading databricks/databricks-dolly-15k ...")
+    ds = load_dataset("databricks/databricks-dolly-15k", split="train")
+
+    conversations = []
+    rng = random.Random(SEED)
+
+    for row in ds:
+        instruction = row["instruction"].strip()
+        context = row.get("context", "").strip()
+        response = row["response"].strip()
+
+        # Skip empties defensively
+        if not instruction or not response:
+            continue
+
+        # Combine instruction and context into the user message
+        if context:
+            user_msg = f"{instruction}\n\n{context}"
+        else:
+            user_msg = instruction
+
+        system_prompt = rng.choice(SYSTEM_PROMPTS)
+        turns = [(user_msg, response)]
+        conversations.append((system_prompt, turns))
+
+    print(f"  Loaded {len(conversations):,} Dolly examples")
+    return conversations
+
+
+# Room for future loaders:
+# def load_ultrachat(): ...
+# def load_oasst1(): ...
+
+
+def _rotate_system_prompts(conversations, seed=SEED):
+    """Assign a random system prompt to each conversation."""
+    rng = random.Random(seed)
+    return [(rng.choice(SYSTEM_PROMPTS), turns) for turns in conversations]
+
+
+def _truncate_turns_to_fit(enc, system_prompt, turns, max_tokens=SEQ_LEN):
+    """
+    Drop trailing turns until the rendered conversation fits in max_tokens.
+
+    Returns trimmed turns, or None if even the first turn alone won't fit.
+    """
+    for n in range(len(turns), 0, -1):
+        trial_turns = turns[:n]
+        rendered, _ = render_conversation(enc, system_prompt, trial_turns)
+        if len(rendered) <= max_tokens:
+            return trial_turns
+    return None
+
+
+def load_ultrachat(split="train_sft", max_conversations=None):
+    """
+    Load UltraChat 200K and extract multi-turn conversations.
+
+    Args:
+        split: 'train_sft' (207K convs) or 'test_sft' (23K, used for val)
+        max_conversations: take at most this many (after loading all). None = all.
+
+    Returns: list of list[(user_msg, assistant_msg)] — raw, no system prompt yet
+    """
+    print(f"Loading HuggingFaceH4/ultrachat_200k split={split} ...")
+    ds = load_dataset("HuggingFaceH4/ultrachat_200k", split=split)
+
+    conversations = []
+    for row in ds:
+        messages = row["messages"]
+        # Walk messages pairwise: [user, assistant, user, assistant, ...]
+        turns = []
+        i = 0
+        while i + 1 < len(messages):
+            u = messages[i]
+            a = messages[i + 1]
+            if u["role"] == "user" and a["role"] == "assistant":
+                turns.append((u["content"].strip(), a["content"].strip()))
+                i += 2
+            else:
+                # Malformed — skip this conversation entirely
+                turns = []
+                break
+        if turns:
+            conversations.append(turns)
+
+    print(f"  Parsed {len(conversations):,} UltraChat conversations from {split}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 3]  # oversample, filters will cut
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+def load_oasst1(max_conversations=None, require_all_rank_zero=True):
+    """
+    Load OASST1 and linearize conversation trees.
+
+    Strategy: for each conversation tree, walk from root to the best leaf.
+    'Best' = the leaf whose path has the lowest sum of ranks (rank 0 = best).
+
+    Args:
+        max_conversations: subsample target
+        require_all_rank_zero: only keep paths where every assistant rank is 0
+
+    Returns: list of list[(user_msg, assistant_msg)]
+    """
+    print("Loading OpenAssistant/oasst1 ...")
+    ds = load_dataset("OpenAssistant/oasst1", split="train")
+
+    # Build message lookup and tree structure
+    print("  Building tree structure ...")
+    messages = {}  # message_id -> row
+    children = {}  # parent_id -> [message_id]
+    roots = []
+
+    for row in ds:
+        mid = row["message_id"]
+        pid = row.get("parent_id")
+        messages[mid] = row
+        if pid is None:
+            roots.append(mid)
+        else:
+            children.setdefault(pid, []).append(mid)
+
+    print(f"  Found {len(messages):,} messages, {len(roots):,} trees")
+
+    # Walk each tree to find best linear path
+    conversations = []
+    drop_reasons = {"non_english_lang": 0, "bad_structure": 0,
+                    "rank_filter": 0, "ok": 0}
+
+    for root_id in roots:
+        root = messages[root_id]
+        # Root must be a prompter message in English
+        if root["role"] != "prompter":
+            drop_reasons["bad_structure"] += 1
+            continue
+        if root.get("lang") != "en":
+            drop_reasons["non_english_lang"] += 1
+            continue
+
+        # Walk greedy best path: at each branch, pick the child with the lowest rank
+        path = [root_id]
+        current = root_id
+        rank_sum = 0
+        bad_rank = False
+
+        while True:
+            kids = children.get(current, [])
+            if not kids:
+                break
+            # For assistant responses, sort by rank ascending (0 is best)
+            # rank can be None for some messages; treat None as worst
+            kids_sorted = sorted(
+                kids,
+                key=lambda mid: (messages[mid].get("rank") if messages[mid].get("rank") is not None else 999)
+            )
+            best_kid_id = kids_sorted[0]
+            best_kid = messages[best_kid_id]
+
+            # Track rank for assistant turns
+            if best_kid["role"] == "assistant":
+                r = best_kid.get("rank")
+                if r is None or r > 0:
+                    bad_rank = True
+                if r is not None:
+                    rank_sum += r
+
+            path.append(best_kid_id)
+            current = best_kid_id
+
+        if require_all_rank_zero and bad_rank:
+            drop_reasons["rank_filter"] += 1
+            continue
+
+        # Convert path to (user, assistant) turns
+        turns = []
+        i = 0
+        structure_ok = True
+        while i + 1 < len(path):
+            u_msg = messages[path[i]]
+            a_msg = messages[path[i + 1]]
+            if u_msg["role"] != "prompter" or a_msg["role"] != "assistant":
+                structure_ok = False
+                break
+            turns.append((u_msg["text"].strip(), a_msg["text"].strip()))
+            i += 2
+
+        if not structure_ok or not turns:
+            drop_reasons["bad_structure"] += 1
+            continue
+
+        conversations.append(turns)
+        drop_reasons["ok"] += 1
+
+    print(f"\n  OASST tree walk results:")
+    for reason, count in drop_reasons.items():
+        print(f"    {reason:>20}: {count:>6}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 2]  # oversample for filters
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+# ─────────────────────────────────────────────────────────────
+# Quality filtering
+# ─────────────────────────────────────────────────────────────
+
+def ascii_ratio(text):
+    """Fraction of characters that are ASCII. Crude English detector."""
+    if not text:
+        return 1.0
+    ascii_count = sum(1 for c in text if ord(c) < 128)
+    return ascii_count / len(text)
+
+
+def is_likely_english(text):
+    """True if a message is probably English (ASCII-ratio based)."""
+    return ascii_ratio(text) >= MIN_ASCII_RATIO
+
+
+def starts_with_refusal_pattern(text):
+    """
+    Check if an assistant response starts with a known AI-disclaimer pattern.
+    Match is on the first ~80 characters, case-insensitive, stripped of leading whitespace.
+    """
+    if not text:
+        return False
+    head = text.strip().lower()[:80]
+    for pat in REFUSAL_PATTERNS:
+        if head.startswith(pat):
+            return True
+    return False
+
+
+def passes_quality_filter(enc, system_prompt, turns, strict=False, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Returns (ok, reason).
+
+    Args:
+        strict: if True, apply production filters (language, absolute assistant
+                minimum length). If False, only apply basic filters (Dolly default).
+        filter_refusals: if True, drop conversations whose assistant response
+                starts with a known "as an AI" / "I am not capable" pattern.
+        max_assistant_tokens: if set, drop conversations whose assistant response
+                exceeds this many tokens in ANY turn. None = no cap.
+    """
+    if len(turns) < 1:
+        return False, "too_few_turns"
+
+    for user_msg, assistant_msg in turns:
+        if not user_msg.strip() or not assistant_msg.strip():
+            return False, "empty_msg"
+
+        # Length check in tokens
+        u_toks = len(encode_plain(enc, user_msg))
+        a_toks = len(encode_plain(enc, assistant_msg))
+        if u_toks < MIN_MSG_TOKENS or a_toks < MIN_MSG_TOKENS:
+            return False, "msg_too_short"
+
+        if strict:
+            # Language filter
+            if not is_likely_english(user_msg) or not is_likely_english(assistant_msg):
+                return False, "non_english"
+
+            # Absolute minimum length on assistant response.
+            # We use an absolute floor rather than a ratio because tying output
+            # length to input length would force the model to pad short
+            # answers to long (e.g. RAG) prompts — teaching it to yap.
+            if a_toks < MIN_ASSISTANT_TOKENS:
+                return False, "assistant_too_short"
+
+        # Optional upper bound (CLI flag). Applies regardless of strict mode
+        # so Dolly could also use it if requested.
+        if max_assistant_tokens is not None and a_toks > max_assistant_tokens:
+            return False, "assistant_too_long"
+
+        if filter_refusals and starts_with_refusal_pattern(assistant_msg):
+            return False, "refusal_pattern"
+
+    return True, "ok"
+
+
+def fits_in_sequence(tokens):
+    """Check that the rendered conversation fits in our seq length."""
+    return len(tokens) <= SEQ_LEN
+
+
+# ─────────────────────────────────────────────────────────────
+# Main pipeline
+# ─────────────────────────────────────────────────────────────
+
+def _filter_and_render(enc, conversations_with_system, strict, cap=None, truncate_to_fit=False, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Apply quality filters and render to (tokens, mask).
+
+    Args:
+        conversations_with_system: list of (system_prompt, turns)
+        strict: use production filters (language, length ratio)
+        cap: stop once we have this many rendered examples (None = no cap)
+        truncate_to_fit: if True, drop trailing turns to fit in SEQ_LEN
+                         instead of dropping the whole conversation
+        filter_refusals: if True, drop conversations with AI-disclaimer openers
+        max_assistant_tokens: if set, drop convs with any assistant msg > this
+
+    Returns: (rendered_list, drop_reasons_dict)
+    """
+    rendered = []
+    drop_reasons = {
+        "too_few_turns": 0, "empty_msg": 0, "msg_too_short": 0,
+        "non_english": 0, "assistant_too_short": 0,
+        "assistant_too_long": 0, "refusal_pattern": 0,
+        "too_long": 0, "ok": 0,
+    }
+
+    for system_prompt, turns in tqdm(conversations_with_system, desc="Rendering"):
+        if cap is not None and len(rendered) >= cap:
+            break
+
+        ok, reason = passes_quality_filter(
+            enc, system_prompt, turns,
+            strict=strict, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+        if not ok:
+            drop_reasons[reason] = drop_reasons.get(reason, 0) + 1
+            continue
+
+        # Try to fit in sequence length
+        if truncate_to_fit:
+            fit_turns = _truncate_turns_to_fit(enc, system_prompt, turns)
+            if fit_turns is None:
+                drop_reasons["too_long"] += 1
+                continue
+            turns_to_render = fit_turns
+        else:
+            turns_to_render = turns
+
+        tokens, mask = render_conversation(enc, system_prompt, turns_to_render)
+        if len(tokens) > SEQ_LEN:
+            drop_reasons["too_long"] += 1
+            continue
+
+        # Must have at least some loss
+        if sum(mask) == 0:
+            drop_reasons["msg_too_short"] += 1
+            continue
+
+        rendered.append((tokens, mask))
+        drop_reasons["ok"] += 1
+
+    return rendered, drop_reasons
+
+
+def _build_train_val_for_dolly(enc, max_assistant_tokens=None):
+    """Dolly: load all, apply basic filters, 95/5 random split."""
+    raw = load_dolly()  # already has system prompts assigned
+    rendered, drops = _filter_and_render(
+        enc, raw, strict=False,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_ultrachat(enc, filter_refusals=False, max_assistant_tokens=None):
+    """UltraChat alone: train_sft for training, test_sft for validation."""
+    train_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+
+    train_convs = _rotate_system_prompts(train_raw, seed=SEED)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+
+    train_rendered, train_drops = _filter_and_render(
+        enc, train_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    return train_rendered, val_rendered, {"train": train_drops, "val": val_drops}
+
+
+def _build_train_val_for_oasst(enc, filter_refusals=False, max_assistant_tokens=None):
+    """OASST alone: load tree, linearize, 95/5 random split."""
+    raw = load_oasst1(max_conversations=OASST_TARGET)
+    convs = _rotate_system_prompts(raw, seed=SEED)
+    rendered, drops = _filter_and_render(
+        enc, convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_prod(enc, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Production: UltraChat train_sft + OASST for training,
+                UltraChat test_sft for validation.
+    """
+    # Training data from both sources
+    print("\n[1/3] Loading UltraChat for training ...")
+    uc_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    uc_convs = _rotate_system_prompts(uc_raw, seed=SEED)
+    uc_rendered, uc_drops = _filter_and_render(
+        enc, uc_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  UltraChat accepted: {len(uc_rendered):,}")
+
+    print("\n[2/3] Loading OASST for training ...")
+    oasst_raw = load_oasst1(max_conversations=OASST_TARGET)
+    oasst_convs = _rotate_system_prompts(oasst_raw, seed=SEED + 2)
+    oasst_rendered, oasst_drops = _filter_and_render(
+        enc, oasst_convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  OASST accepted: {len(oasst_rendered):,}")
+
+    # Combine and shuffle
+    train_rendered = uc_rendered + oasst_rendered
+    rng = random.Random(SEED)
+    rng.shuffle(train_rendered)
+    print(f"\n  Combined training set: {len(train_rendered):,} conversations")
+
+    # Validation from UltraChat test_sft only (clean, no OASST noise)
+    print("\n[3/3] Loading UltraChat test_sft for validation ...")
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  Val accepted: {len(val_rendered):,}")
+
+    return train_rendered, val_rendered, {
+        "ultrachat_train": uc_drops,
+        "oasst_train": oasst_drops,
+        "val": val_drops,
+        "sources": {"ultrachat": len(uc_rendered), "oasst": len(oasst_rendered)},
+    }
+
+
+def process_dataset(dataset_name, output_dir, filter_refusals=False, max_assistant_tokens=None):
+    enc = get_tokenizer()
+
+    # Route to the appropriate builder
+    print(f"\n{'='*70}")
+    print(f"Processing dataset: {dataset_name}")
+    print(f"  filter_refusals:        {filter_refusals}")
+    print(f"  max_assistant_tokens:   {max_assistant_tokens}")
+    print(f"{'='*70}")
+
+    if dataset_name == "dolly":
+        # Dolly: refusal filter not wired (single-turn, rare patterns there)
+        train_data, val_data, filter_info = _build_train_val_for_dolly(
+            enc, max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "ultrachat":
+        train_data, val_data, filter_info = _build_train_val_for_ultrachat(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "oasst":
+        train_data, val_data, filter_info = _build_train_val_for_oasst(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "prod":
+        train_data, val_data, filter_info = _build_train_val_for_prod(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    else:
+        raise ValueError(f"Unknown dataset: {dataset_name}")
+
+    if len(train_data) == 0:
+        raise RuntimeError("No training examples survived filtering. Check your data.")
+    if len(val_data) == 0:
+        raise RuntimeError("No validation examples survived filtering.")
+
+    print(f"\n  Train: {len(train_data):,} examples")
+    print(f"  Val:   {len(val_data):,} examples")
+
+    # Pad and convert to arrays
+    print("\nPadding and converting to arrays ...")
+
+    def to_arrays(data):
+        n = len(data)
+        tokens_arr = np.full((n, SEQ_LEN), PAD_TOKEN_ID, dtype=np.uint32)
+        mask_arr = np.zeros((n, SEQ_LEN), dtype=np.uint8)
+        for i, (toks, msk) in enumerate(data):
+            toks_padded, msk_padded = pad_to_seq_len(toks, msk)
+            tokens_arr[i] = toks_padded
+            mask_arr[i] = msk_padded
+        return tokens_arr, mask_arr
+
+    train_tokens, train_mask = to_arrays(train_data)
+    val_tokens, val_mask = to_arrays(val_data)
+
+    # Save
+    os.makedirs(output_dir, exist_ok=True)
+    np.save(os.path.join(output_dir, "sft_train.npy"), train_tokens)
+    np.save(os.path.join(output_dir, "sft_train_mask.npy"), train_mask)
+    np.save(os.path.join(output_dir, "sft_val.npy"), val_tokens)
+    np.save(os.path.join(output_dir, "sft_val_mask.npy"), val_mask)
+
+    # Metadata
+    def stats(tokens_arr, mask_arr):
+        real_lens = (tokens_arr != PAD_TOKEN_ID).sum(axis=1)
+        loss_fractions = mask_arr.sum(axis=1) / real_lens.clip(min=1)
+        return {
+            "num_examples": int(tokens_arr.shape[0]),
+            "total_tokens": int(real_lens.sum()),
+            "total_loss_tokens": int(mask_arr.sum()),
+            "avg_length": float(real_lens.mean()),
+            "median_length": float(np.median(real_lens)),
+            "min_length": int(real_lens.min()),
+            "max_length": int(real_lens.max()),
+            "avg_loss_fraction": float(loss_fractions.mean()),
+        }
+
+    metadata = {
+        "dataset": dataset_name,
+        "seq_len": SEQ_LEN,
+        "pad_token_id": PAD_TOKEN_ID,
+        "im_start_id": IM_START_ID,
+        "im_end_id": IM_END_ID,
+        "eot_id": EOT_ID,
+        "system_prompts": SYSTEM_PROMPTS,
+        "filter_refusals": filter_refusals,
+        "refusal_patterns": REFUSAL_PATTERNS if filter_refusals else None,
+        "max_assistant_tokens": max_assistant_tokens,
+        "filter_results": filter_info,
+        "train": stats(train_tokens, train_mask),
+        "val": stats(val_tokens, val_mask),
+    }
+
+    with open(os.path.join(output_dir, "sft_metadata.json"), "w") as f:
+        json.dump(metadata, f, indent=2)
+
+    print(f"\n  Wrote arrays and metadata to {output_dir}/")
+
+    # Preview
+    print("\n" + "=" * 70)
+    print("PREVIEW: 3 random examples (abbreviated)")
+    print("=" * 70)
+    rng = random.Random(SEED)
+    preview_indices = rng.sample(range(len(train_data)), k=min(3, len(train_data)))
+    for idx in preview_indices:
+        tokens, mask = train_data[idx]
+        preview_example(enc, tokens, mask, max_tokens=80)
+
+    print("\nDone. Next: run test_preprocessing.py to verify correctness.")
+    return metadata
+
+
+def preview_example(enc, tokens, mask, max_tokens=80):
+    """Print a human-readable preview showing tokens with their mask values."""
+    print("\n" + "-" * 70)
+    print(f"Total tokens: {len(tokens)}, loss tokens: {sum(mask)}")
+    print(f"First {min(max_tokens, len(tokens))} tokens:")
+    print(f"{'IDX':>4} {'TOK_ID':>7} {'MASK':>4}  TEXT")
+    for i, (tok, m) in enumerate(zip(tokens[:max_tokens], mask[:max_tokens])):
+        try:
+            text = enc.decode([tok])
+        except Exception:
+            text = "<decode-error>"
+        marker = "◀LOSS" if m else ""
+        # Escape newlines for readability
+        text_display = repr(text)[1:-1][:40]
+        print(f"{i:>4} {tok:>7} {m:>4}  {text_display} {marker}")
+    if len(tokens) > max_tokens:
+        print(f"  ... ({len(tokens) - max_tokens} more tokens)")
+
+
+# ─────────────────────────────────────────────────────────────
+# CLI entry point
+# ─────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--dataset",
+        choices=["dolly", "ultrachat", "oasst", "prod"],
+        default="dolly",
+        help="Which dataset to process. 'prod' = UltraChat + OASST mix (recommended for real run)."
+    )
+    parser.add_argument(
+        "--output", type=str, default=None,
+        help="Output directory (default: sft_data_<dataset>[_clean][_maxN])"
+    )
+    parser.add_argument(
+        "--filter-refusals", action="store_true",
+        help="Drop conversations whose assistant response starts with 'as an AI', "
+             "'I am not capable', etc. Does not apply to Dolly."
+    )
+    parser.add_argument(
+        "--max-assistant-tokens", type=int, default=None,
+        help="If set, drop conversations where any assistant message exceeds "
+             "this many tokens. Biases training toward shorter responses. "
+             "Try 700 (drops ~50%% UltraChat, teaches concise style) or "
+             "1000 (drops ~30%%, mild outlier filter)."
+    )
+    args = parser.parse_args()
+
+    # Default output dir reflects what filters were applied so runs don't collide.
+    if args.output:
+        output_dir = args.output
+    else:
+        suffix = ""
+        if args.filter_refusals:
+            suffix += "_clean"
+        if args.max_assistant_tokens is not None:
+            suffix += f"_max{args.max_assistant_tokens}"
+        output_dir = f"sft_data_{args.dataset}{suffix}"
+
+    process_dataset(
+        args.dataset, output_dir,
+        filter_refusals=args.filter_refusals,
+        max_assistant_tokens=args.max_assistant_tokens,
+    )

workspace/Alter_Ego/prep_sft2.old

@@ -0,0 +1,998 @@
+"""
+prep_sft_data.py
+
+Tokenizes chat datasets for SFT training. Produces paired (tokens, mask) arrays
+where mask=1 on tokens the model should learn to generate and mask=0 elsewhere.
+
+Supported datasets:
+- dolly     — Dolly 15K (databricks/databricks-dolly-15k), single-turn
+- ultrachat — UltraChat 200K (HuggingFaceH4/ultrachat_200k), multi-turn
+- oasst     — OpenAssistant OASST1 (OpenAssistant/oasst1), multi-turn tree
+- prod      — UltraChat + OASST mix (~40M tokens, production run)
+
+Output files (written to OUTPUT_DIR):
+    sft_train.npy          uint32 (N, T)   token IDs
+    sft_train_mask.npy     uint8  (N, T)   1 = loss on this token
+    sft_val.npy            uint32 (M, T)
+    sft_val_mask.npy       uint8  (M, T)
+    sft_metadata.json                      statistics and provenance
+
+Run:
+    python prep_sft_data.py --dataset dolly --output sft_data_dolly
+    python prep_sft_data.py --dataset prod --output sft_data_prod
+"""
+
+import argparse
+import json
+import os
+import random
+from pathlib import Path
+
+import numpy as np
+import tiktoken
+from datasets import load_dataset  # pip install datasets
+from tqdm import tqdm  # pip install tqdm
+
+
+# ─────────────────────────────────────────────────────────────
+# Constants — must match SFT_PLAN.md
+# ─────────────────────────────────────────────────────────────
+
+SEQ_LEN = 2048            # T — matches pretraining context
+VAL_FRACTION = 0.05       # 5% held out for eval loss
+SEED = 42                 # reproducibility
+
+# Special token IDs — see SFT_PLAN.md §2.1
+EOT_ID = 100257           # <|endoftext|> (native cl100k_base)
+IM_START_ID = 100277      # <|im_start|> (NEW, added for SFT)
+IM_END_ID = 100278        # <|im_end|> (NEW, added for SFT)
+
+# Padding uses <|endoftext|> — NOT token 0.
+# Token 0 in cl100k_base is '!' which appears frequently in real chat text.
+# Using <|endoftext|> ensures "is this position padding?" has one unambiguous answer.
+PAD_TOKEN_ID = EOT_ID     # = 100257
+
+# Quality filters
+MIN_MSG_TOKENS = 5        # drop conversations with any msg shorter than this
+MIN_TURNS = 2             # need at least one user + one assistant
+
+# New filters for production datasets (see SFT_PLAN.md §4.3)
+MIN_ASCII_RATIO = 0.80    # drop non-English content (crude but dependency-free)
+MIN_ASSISTANT_TOKENS = 10 # assistant response must be at least this many tokens.
+                          # Absolute floor, not a ratio — we DON'T want to tie
+                          # output length to input length (would force verbose
+                          # responses to long RAG prompts).
+# Optional upper bound on assistant response length.
+# Enabled via --max-assistant-tokens N on the CLI. Default None = no cap.
+# Recommended values when enabled:
+#   700  -> drops ~50% of UltraChat (essays), teaches concise responses
+#   1000 -> drops ~30% of UltraChat (outliers only)
+# Setting a cap trades dataset size for response-length bias during training.
+MAX_ASSISTANT_TOKENS_DEFAULT = None
+
+# Production dataset target sizes (see SFT_PLAN.md §4.2)
+ULTRACHAT_TARGET = 25_714    # ~36M tokens at avg 1,400 tok/conv
+OASST_TARGET = 5_000         # ~4M tokens at avg 800 tok/conv
+
+# Refusal / "as an AI" patterns — optional filter.
+# Matches if any assistant response STARTS WITH (case-insensitive) one of these.
+# Kept strict (first-word/phrase only) to avoid false positives on legit
+# discussions about AI.
+REFUSAL_PATTERNS = [
+    "as an ai",
+    "as a language model",
+    "as an ai language model",
+    "as an ai assistant",
+    "i am an ai",
+    "i'm an ai",
+    "i am just an ai",
+    "i'm just an ai",
+    "i am not capable",
+    "i'm not capable",
+    "i am not able",
+    "i'm not able",
+    "i don't have the ability",
+    "i do not have the ability",
+    "i don't have access",
+    "i do not have access",
+    "i am unable to",
+    "i'm unable to",
+    "i cannot browse",
+    "i can't browse",
+    "i cannot provide",
+    "i can't provide personal",
+    "i don't have personal",
+    "i do not have personal",
+    "i don't have feelings",
+    "i do not have feelings",
+    "i don't have emotions",
+    "i do not have emotions",
+    "i don't have opinions",
+    "i do not have opinions",
+    "as a responsible ai",
+    "as an artificial intelligence",
+]
+
+# Varied system prompts — one is chosen per conversation (SFT_PLAN.md §3)
+SYSTEM_PROMPTS = [
+    # Core - smart, casual, engaged (40%)
+    "You are Alter Ego. You enjoy explaining things clearly. Speak casually and use contractions.",
+    "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out.",
+    "You are Alter Ego. Be friendly and get to the point. Keep it natural.",
+    "You are Alter Ego. Explain things simply and conversationally.",
+
+    # Warmer / approachable (30%)
+    "You are Alter Ego. Be warm, relaxed, and conversational.",
+    "You are Alter Ego. Share what you know in a friendly, easy-to-understand way.",
+    "You are Alter Ego. You're a clever and approachable assistant. Keep it casual.",
+
+    # Bridge to standard helpful (20%)
+    "You are Alter Ego, a helpful and friendly AI.",
+    "You are Alter Ego. Provide clear, accurate answers.",
+
+    # Nerdy-adjacent (10%)
+    "You are Alter Ego. You find learning interesting and enjoy breaking down complex topics into simple terms.",
+]
+
+
+# ─────────────────────────────────────────────────────────────
+# Extended tokenizer with ChatML special tokens
+# ─────────────────────────────────────────────────────────────
+
+def get_tokenizer():
+    """
+    Returns cl100k_base extended with <|im_start|> and <|im_end|>.
+
+    Uses fixed IDs 100277 and 100278 so that the same function can be called
+    during prep, training, and inference without surprises.
+    """
+    base = tiktoken.get_encoding("cl100k_base")
+    enc = tiktoken.Encoding(
+        name="cl100k_alterego",
+        pat_str=base._pat_str,
+        mergeable_ranks=base._mergeable_ranks,
+        special_tokens={
+            **base._special_tokens,
+            "<|im_start|>": IM_START_ID,
+            "<|im_end|>": IM_END_ID,
+        },
+    )
+    return enc
+
+
+# ─────────────────────────────────────────────────────────────
+# Chat template rendering
+# ─────────────────────────────────────────────────────────────
+
+def encode_plain(enc, text):
+    """Encode ordinary content — no special tokens allowed in user data."""
+    return enc.encode(text, allowed_special=set(), disallowed_special=())
+
+
+def encode_controls(enc, text):
+    """
+    Encode a string that contains our control tokens (<|im_start|>, <|im_end|>).
+
+    Only called on strings WE construct — never on user/dataset content.
+    """
+    return enc.encode(
+        text,
+        allowed_special={"<|im_start|>", "<|im_end|>"},
+        disallowed_special=(),
+    )
+
+
+def render_conversation(enc, system_prompt, turns, max_len=SEQ_LEN):
+    """
+    Render a full conversation to (tokens, mask) arrays.
+
+    Args:
+        enc: the extended tiktoken encoding
+        system_prompt: str, the system message content
+        turns: list of (user_msg, assistant_msg) tuples
+        max_len: truncate to this many tokens if necessary
+
+    Returns:
+        tokens: list[int], length <= max_len (not yet padded)
+        mask:   list[int], same length
+                mask[i] = 1 if we train on predicting tokens[i], 0 otherwise
+
+    Loss mask rules (see SFT_PLAN.md §4.5):
+        - System turn: all masked (0)
+        - User turn: all masked (0)
+        - Assistant prefix (<|im_start|>assistant\\n): masked (0)
+          — the trainer provides this; model shouldn't be penalized for it
+        - Assistant content + <|im_end|>: LOSS (1)
+          — model must learn to generate content AND stop
+    """
+    tokens = []
+    mask = []
+
+    def append(toks, loss):
+        tokens.extend(toks)
+        mask.extend([loss] * len(toks))
+
+    # ---- System turn (no loss)
+    system_block = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
+    append(encode_controls(enc, "<|im_start|>system\n"), 0)
+    append(encode_plain(enc, system_prompt), 0)
+    append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+    # ---- Turns
+    for user_msg, assistant_msg in turns:
+        # User turn — no loss on any part
+        append(encode_controls(enc, "<|im_start|>user\n"), 0)
+        append(encode_plain(enc, user_msg), 0)
+        append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+        # Assistant turn — prefix is masked, content + <|im_end|> gets loss
+        append(encode_controls(enc, "<|im_start|>assistant\n"), 0)
+        append(encode_plain(enc, assistant_msg), 1)
+        # The <|im_end|> after assistant content IS part of the loss
+        # so the model learns to terminate its turn.
+        append(encode_controls(enc, "<|im_end|>"), 1)
+        # The trailing newline after <|im_end|> (between turns) is masked
+        # — it's structural, not content.
+        append(encode_controls(enc, "\n"), 0)
+
+    # Truncate if too long (rare; we pre-filter but defensive here)
+    if len(tokens) > max_len:
+        tokens = tokens[:max_len]
+        mask = mask[:max_len]
+
+    assert len(tokens) == len(mask), "Token/mask length mismatch"
+    return tokens, mask
+
+
+def pad_to_seq_len(tokens, mask, target_len=SEQ_LEN):
+    """Pad to fixed length. Padding has mask=0."""
+    assert len(tokens) <= target_len
+    pad_needed = target_len - len(tokens)
+    tokens = tokens + [PAD_TOKEN_ID] * pad_needed
+    mask = mask + [0] * pad_needed
+    return tokens, mask
+
+
+# ─────────────────────────────────────────────────────────────
+# Dataset loaders
+# ─────────────────────────────────────────────────────────────
+
+def load_dolly():
+    """
+    Load Dolly 15K and normalize to a list of (system_prompt, turns) tuples.
+
+    Dolly has single-turn instruction/context/response triples. We convert
+    to one-turn conversations with random system prompts.
+
+    Returns: list of (system_prompt, [(user_msg, assistant_msg)])
+    """
+    print("Loading databricks/databricks-dolly-15k ...")
+    ds = load_dataset("databricks/databricks-dolly-15k", split="train")
+
+    conversations = []
+    rng = random.Random(SEED)
+
+    for row in ds:
+        instruction = row["instruction"].strip()
+        context = row.get("context", "").strip()
+        response = row["response"].strip()
+
+        # Skip empties defensively
+        if not instruction or not response:
+            continue
+
+        # Combine instruction and context into the user message
+        if context:
+            user_msg = f"{instruction}\n\n{context}"
+        else:
+            user_msg = instruction
+
+        system_prompt = rng.choice(SYSTEM_PROMPTS)
+        turns = [(user_msg, response)]
+        conversations.append((system_prompt, turns))
+
+    print(f"  Loaded {len(conversations):,} Dolly examples")
+    return conversations
+
+
+# Room for future loaders:
+# def load_ultrachat(): ...
+# def load_oasst1(): ...
+
+
+def load_synthetic_chat(path="alter_ego_synthetic_clean.jsonl", repeat=5):
+    """
+    Load Gemini-generated synthetic conversation pairs.
+
+    These pairs (greetings, identity, capabilities, etc.) are rare in
+    UltraChat/OASST so we repeat them multiple times to give them weight
+    in the training mix.
+
+    Args:
+        path: JSONL file with rows {category, user, assistant}
+        repeat: how many copies of each pair to add to the training set.
+                Higher = stronger learning of these patterns. Default 5
+                means each pair is seen 5 times during 1 epoch.
+
+    Returns: list of [(user_msg, assistant_msg)] — single-turn conversations
+    """
+    if not os.path.isfile(path):
+        print(f"  Synthetic data file not found at {path} — skipping.")
+        return []
+
+    print(f"Loading synthetic chat data from {path} ...")
+    rows = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            row = json.loads(line)
+            rows.append(row)
+
+    print(f"  Loaded {len(rows)} unique synthetic pairs")
+    print(f"  Replicating {repeat}x for training weight = {len(rows) * repeat:,} examples")
+
+    # Convert to (system_prompt, turns) format. We use a NEUTRAL system prompt
+    # for these so the model learns the persona is intrinsic, not prompt-dependent.
+    # We don't randomize over SYSTEM_PROMPTS here — these pairs deserve to be
+    # paired with simple, consistent framing so the model learns "this is just
+    # how Alter Ego talks."
+    conversations = []
+    rng = random.Random(SEED + 100)
+    for row in rows:
+        # Mix of system prompts: half generic ("You are Alter Ego."), half
+        # from our normal pool. This lets the model generalize across system
+        # prompt variations while strongly anchoring the basic identity.
+        for _ in range(repeat):
+            if rng.random() < 0.5:
+                sys_prompt = "You are Alter Ego."
+            else:
+                sys_prompt = rng.choice(SYSTEM_PROMPTS)
+            turns = [(row["user"], row["assistant"])]
+            conversations.append((sys_prompt, turns))
+
+    return conversations
+
+
+def _rotate_system_prompts(conversations, seed=SEED):
+    """Assign a random system prompt to each conversation."""
+    rng = random.Random(seed)
+    return [(rng.choice(SYSTEM_PROMPTS), turns) for turns in conversations]
+
+
+def _truncate_turns_to_fit(enc, system_prompt, turns, max_tokens=SEQ_LEN):
+    """
+    Drop trailing turns until the rendered conversation fits in max_tokens.
+
+    Returns trimmed turns, or None if even the first turn alone won't fit.
+    """
+    for n in range(len(turns), 0, -1):
+        trial_turns = turns[:n]
+        rendered, _ = render_conversation(enc, system_prompt, trial_turns)
+        if len(rendered) <= max_tokens:
+            return trial_turns
+    return None
+
+
+def load_ultrachat(split="train_sft", max_conversations=None):
+    """
+    Load UltraChat 200K and extract multi-turn conversations.
+
+    Args:
+        split: 'train_sft' (207K convs) or 'test_sft' (23K, used for val)
+        max_conversations: take at most this many (after loading all). None = all.
+
+    Returns: list of list[(user_msg, assistant_msg)] — raw, no system prompt yet
+    """
+    print(f"Loading HuggingFaceH4/ultrachat_200k split={split} ...")
+    ds = load_dataset("HuggingFaceH4/ultrachat_200k", split=split)
+
+    conversations = []
+    for row in ds:
+        messages = row["messages"]
+        # Walk messages pairwise: [user, assistant, user, assistant, ...]
+        turns = []
+        i = 0
+        while i + 1 < len(messages):
+            u = messages[i]
+            a = messages[i + 1]
+            if u["role"] == "user" and a["role"] == "assistant":
+                turns.append((u["content"].strip(), a["content"].strip()))
+                i += 2
+            else:
+                # Malformed — skip this conversation entirely
+                turns = []
+                break
+        if turns:
+            conversations.append(turns)
+
+    print(f"  Parsed {len(conversations):,} UltraChat conversations from {split}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 3]  # oversample, filters will cut
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+def load_oasst1(max_conversations=None, require_all_rank_zero=True):
+    """
+    Load OASST1 and linearize conversation trees.
+
+    Strategy: for each conversation tree, walk from root to the best leaf.
+    'Best' = the leaf whose path has the lowest sum of ranks (rank 0 = best).
+
+    Args:
+        max_conversations: subsample target
+        require_all_rank_zero: only keep paths where every assistant rank is 0
+
+    Returns: list of list[(user_msg, assistant_msg)]
+    """
+    print("Loading OpenAssistant/oasst1 ...")
+    ds = load_dataset("OpenAssistant/oasst1", split="train")
+
+    # Build message lookup and tree structure
+    print("  Building tree structure ...")
+    messages = {}  # message_id -> row
+    children = {}  # parent_id -> [message_id]
+    roots = []
+
+    for row in ds:
+        mid = row["message_id"]
+        pid = row.get("parent_id")
+        messages[mid] = row
+        if pid is None:
+            roots.append(mid)
+        else:
+            children.setdefault(pid, []).append(mid)
+
+    print(f"  Found {len(messages):,} messages, {len(roots):,} trees")
+
+    # Walk each tree to find best linear path
+    conversations = []
+    drop_reasons = {"non_english_lang": 0, "bad_structure": 0,
+                    "rank_filter": 0, "ok": 0}
+
+    for root_id in roots:
+        root = messages[root_id]
+        # Root must be a prompter message in English
+        if root["role"] != "prompter":
+            drop_reasons["bad_structure"] += 1
+            continue
+        if root.get("lang") != "en":
+            drop_reasons["non_english_lang"] += 1
+            continue
+
+        # Walk greedy best path: at each branch, pick the child with the lowest rank
+        path = [root_id]
+        current = root_id
+        rank_sum = 0
+        bad_rank = False
+
+        while True:
+            kids = children.get(current, [])
+            if not kids:
+                break
+            # For assistant responses, sort by rank ascending (0 is best)
+            # rank can be None for some messages; treat None as worst
+            kids_sorted = sorted(
+                kids,
+                key=lambda mid: (messages[mid].get("rank") if messages[mid].get("rank") is not None else 999)
+            )
+            best_kid_id = kids_sorted[0]
+            best_kid = messages[best_kid_id]
+
+            # Track rank for assistant turns
+            if best_kid["role"] == "assistant":
+                r = best_kid.get("rank")
+                if r is None or r > 0:
+                    bad_rank = True
+                if r is not None:
+                    rank_sum += r
+
+            path.append(best_kid_id)
+            current = best_kid_id
+
+        if require_all_rank_zero and bad_rank:
+            drop_reasons["rank_filter"] += 1
+            continue
+
+        # Convert path to (user, assistant) turns
+        turns = []
+        i = 0
+        structure_ok = True
+        while i + 1 < len(path):
+            u_msg = messages[path[i]]
+            a_msg = messages[path[i + 1]]
+            if u_msg["role"] != "prompter" or a_msg["role"] != "assistant":
+                structure_ok = False
+                break
+            turns.append((u_msg["text"].strip(), a_msg["text"].strip()))
+            i += 2
+
+        if not structure_ok or not turns:
+            drop_reasons["bad_structure"] += 1
+            continue
+
+        conversations.append(turns)
+        drop_reasons["ok"] += 1
+
+    print(f"\n  OASST tree walk results:")
+    for reason, count in drop_reasons.items():
+        print(f"    {reason:>20}: {count:>6}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 2]  # oversample for filters
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+# ─────────────────────────────────────────────────────────────
+# Quality filtering
+# ─────────────────────────────────────────────────────────────
+
+def ascii_ratio(text):
+    """Fraction of characters that are ASCII. Crude English detector."""
+    if not text:
+        return 1.0
+    ascii_count = sum(1 for c in text if ord(c) < 128)
+    return ascii_count / len(text)
+
+
+def is_likely_english(text):
+    """True if a message is probably English (ASCII-ratio based)."""
+    return ascii_ratio(text) >= MIN_ASCII_RATIO
+
+
+def starts_with_refusal_pattern(text):
+    """
+    Check if an assistant response starts with a known AI-disclaimer pattern.
+    Match is on the first ~80 characters, case-insensitive, stripped of leading whitespace.
+    """
+    if not text:
+        return False
+    head = text.strip().lower()[:80]
+    for pat in REFUSAL_PATTERNS:
+        if head.startswith(pat):
+            return True
+    return False
+
+
+def passes_quality_filter(enc, system_prompt, turns, strict=False, filter_refusals=False, max_assistant_tokens=None, bypass_min_tokens=False):
+    """
+    Returns (ok, reason).
+
+    Args:
+        strict: if True, apply production filters (language, absolute assistant
+                minimum length). If False, only apply basic filters (Dolly default).
+        filter_refusals: if True, drop conversations whose assistant response
+                starts with a known "as an AI" / "I am not capable" pattern.
+        max_assistant_tokens: if set, drop conversations whose assistant response
+                exceeds this many tokens in ANY turn. None = no cap.
+        bypass_min_tokens: if True, skip the MIN_MSG_TOKENS check. Use ONLY for
+                hand-curated synthetic data where short messages (e.g. 'hi',
+                'thanks') are intentional and valid.
+    """
+    if len(turns) < 1:
+        return False, "too_few_turns"
+
+    for user_msg, assistant_msg in turns:
+        if not user_msg.strip() or not assistant_msg.strip():
+            return False, "empty_msg"
+
+        # Length check in tokens
+        u_toks = len(encode_plain(enc, user_msg))
+        a_toks = len(encode_plain(enc, assistant_msg))
+        if not bypass_min_tokens:
+            if u_toks < MIN_MSG_TOKENS or a_toks < MIN_MSG_TOKENS:
+                return False, "msg_too_short"
+
+        if strict:
+            # Language filter
+            if not is_likely_english(user_msg) or not is_likely_english(assistant_msg):
+                return False, "non_english"
+
+            # Absolute minimum length on assistant response.
+            # We use an absolute floor rather than a ratio because tying output
+            # length to input length would force the model to pad short
+            # answers to long (e.g. RAG) prompts — teaching it to yap.
+            if a_toks < MIN_ASSISTANT_TOKENS:
+                return False, "assistant_too_short"
+
+        # Optional upper bound (CLI flag). Applies regardless of strict mode
+        # so Dolly could also use it if requested.
+        if max_assistant_tokens is not None and a_toks > max_assistant_tokens:
+            return False, "assistant_too_long"
+
+        if filter_refusals and starts_with_refusal_pattern(assistant_msg):
+            return False, "refusal_pattern"
+
+    return True, "ok"
+
+
+def fits_in_sequence(tokens):
+    """Check that the rendered conversation fits in our seq length."""
+    return len(tokens) <= SEQ_LEN
+
+
+# ─────────────────────────────────────────────────────────────
+# Main pipeline
+# ─────────────────────────────────────────────────────────────
+
+def _filter_and_render(enc, conversations_with_system, strict, cap=None, truncate_to_fit=False, filter_refusals=False, max_assistant_tokens=None, bypass_min_tokens=False):
+    """
+    Apply quality filters and render to (tokens, mask).
+
+    Args:
+        conversations_with_system: list of (system_prompt, turns)
+        strict: use production filters (language, length ratio)
+        cap: stop once we have this many rendered examples (None = no cap)
+        truncate_to_fit: if True, drop trailing turns to fit in SEQ_LEN
+                         instead of dropping the whole conversation
+        filter_refusals: if True, drop conversations with AI-disclaimer openers
+        max_assistant_tokens: if set, drop convs with any assistant msg > this
+        bypass_min_tokens: skip MIN_MSG_TOKENS check (for hand-curated synthetic
+                           data where short messages are intentional)
+
+    Returns: (rendered_list, drop_reasons_dict)
+    """
+    rendered = []
+    drop_reasons = {
+        "too_few_turns": 0, "empty_msg": 0, "msg_too_short": 0,
+        "non_english": 0, "assistant_too_short": 0,
+        "assistant_too_long": 0, "refusal_pattern": 0,
+        "too_long": 0, "ok": 0,
+    }
+
+    for system_prompt, turns in tqdm(conversations_with_system, desc="Rendering"):
+        if cap is not None and len(rendered) >= cap:
+            break
+
+        ok, reason = passes_quality_filter(
+            enc, system_prompt, turns,
+            strict=strict, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+            bypass_min_tokens=bypass_min_tokens,
+        )
+        if not ok:
+            drop_reasons[reason] = drop_reasons.get(reason, 0) + 1
+            continue
+
+        # Try to fit in sequence length
+        if truncate_to_fit:
+            fit_turns = _truncate_turns_to_fit(enc, system_prompt, turns)
+            if fit_turns is None:
+                drop_reasons["too_long"] += 1
+                continue
+            turns_to_render = fit_turns
+        else:
+            turns_to_render = turns
+
+        tokens, mask = render_conversation(enc, system_prompt, turns_to_render)
+        if len(tokens) > SEQ_LEN:
+            drop_reasons["too_long"] += 1
+            continue
+
+        # Must have at least some loss
+        if sum(mask) == 0:
+            drop_reasons["msg_too_short"] += 1
+            continue
+
+        rendered.append((tokens, mask))
+        drop_reasons["ok"] += 1
+
+    return rendered, drop_reasons
+
+
+def _build_train_val_for_dolly(enc, max_assistant_tokens=None):
+    """Dolly: load all, apply basic filters, 95/5 random split."""
+    raw = load_dolly()  # already has system prompts assigned
+    rendered, drops = _filter_and_render(
+        enc, raw, strict=False,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_ultrachat(enc, filter_refusals=False, max_assistant_tokens=None):
+    """UltraChat alone: train_sft for training, test_sft for validation."""
+    train_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+
+    train_convs = _rotate_system_prompts(train_raw, seed=SEED)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+
+    train_rendered, train_drops = _filter_and_render(
+        enc, train_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    return train_rendered, val_rendered, {"train": train_drops, "val": val_drops}
+
+
+def _build_train_val_for_oasst(enc, filter_refusals=False, max_assistant_tokens=None):
+    """OASST alone: load tree, linearize, 95/5 random split."""
+    raw = load_oasst1(max_conversations=OASST_TARGET)
+    convs = _rotate_system_prompts(raw, seed=SEED)
+    rendered, drops = _filter_and_render(
+        enc, convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_prod(enc, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Production: UltraChat train_sft + OASST for training,
+                UltraChat test_sft for validation.
+    """
+    # Training data from both sources
+    print("\n[1/3] Loading UltraChat for training ...")
+    uc_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    uc_convs = _rotate_system_prompts(uc_raw, seed=SEED)
+    uc_rendered, uc_drops = _filter_and_render(
+        enc, uc_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  UltraChat accepted: {len(uc_rendered):,}")
+
+    print("\n[2/3] Loading OASST for training ...")
+    oasst_raw = load_oasst1(max_conversations=OASST_TARGET)
+    oasst_convs = _rotate_system_prompts(oasst_raw, seed=SEED + 2)
+    oasst_rendered, oasst_drops = _filter_and_render(
+        enc, oasst_convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  OASST accepted: {len(oasst_rendered):,}")
+
+    # ── Synthetic conversational pairs (greetings, identity, capabilities) ──
+    # These are ABSENT from UltraChat/OASST so we add them explicitly. They
+    # train the model to handle short casual inputs and to know its identity.
+    # Loaded only if alter_ego_synthetic_clean.jsonl exists.
+    print("\n[2.5/3] Loading synthetic conversational pairs ...")
+    synth_raw = load_synthetic_chat(repeat=8)  # each pair sees model 8 times
+    synth_drops = {"ok": 0}
+    synth_rendered = []
+    if synth_raw:
+        # Synthetic pairs use 'strict=False' since they're already curated by
+        # us and don't need the language/refusal/length-floor checks.
+        # bypass_min_tokens=True because short greetings ("hi", "thanks") are
+        # intentional and would otherwise be dropped by MIN_MSG_TOKENS=5.
+        synth_rendered, synth_drops = _filter_and_render(
+            enc, synth_raw, strict=False, truncate_to_fit=True,
+            filter_refusals=False,            # we wrote these, no refusals
+            max_assistant_tokens=None,        # already short by design
+            bypass_min_tokens=True,           # CRITICAL: keep "hi", "yo", etc.
+        )
+        print(f"  Synthetic accepted: {len(synth_rendered):,}")
+
+    # Combine and shuffle
+    train_rendered = uc_rendered + oasst_rendered + synth_rendered
+    rng = random.Random(SEED)
+    rng.shuffle(train_rendered)
+    print(f"\n  Combined training set: {len(train_rendered):,} conversations")
+
+    # Validation from UltraChat test_sft only (clean, no OASST noise)
+    print("\n[3/3] Loading UltraChat test_sft for validation ...")
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  Val accepted: {len(val_rendered):,}")
+
+    return train_rendered, val_rendered, {
+        "ultrachat_train": uc_drops,
+        "oasst_train": oasst_drops,
+        "synthetic_train": synth_drops,
+        "val": val_drops,
+        "sources": {
+            "ultrachat": len(uc_rendered),
+            "oasst": len(oasst_rendered),
+            "synthetic": len(synth_rendered),
+        },
+    }
+
+
+def process_dataset(dataset_name, output_dir, filter_refusals=False, max_assistant_tokens=None):
+    enc = get_tokenizer()
+
+    # Route to the appropriate builder
+    print(f"\n{'='*70}")
+    print(f"Processing dataset: {dataset_name}")
+    print(f"  filter_refusals:        {filter_refusals}")
+    print(f"  max_assistant_tokens:   {max_assistant_tokens}")
+    print(f"{'='*70}")
+
+    if dataset_name == "dolly":
+        # Dolly: refusal filter not wired (single-turn, rare patterns there)
+        train_data, val_data, filter_info = _build_train_val_for_dolly(
+            enc, max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "ultrachat":
+        train_data, val_data, filter_info = _build_train_val_for_ultrachat(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "oasst":
+        train_data, val_data, filter_info = _build_train_val_for_oasst(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "prod":
+        train_data, val_data, filter_info = _build_train_val_for_prod(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    else:
+        raise ValueError(f"Unknown dataset: {dataset_name}")
+
+    if len(train_data) == 0:
+        raise RuntimeError("No training examples survived filtering. Check your data.")
+    if len(val_data) == 0:
+        raise RuntimeError("No validation examples survived filtering.")
+
+    print(f"\n  Train: {len(train_data):,} examples")
+    print(f"  Val:   {len(val_data):,} examples")
+
+    # Pad and convert to arrays
+    print("\nPadding and converting to arrays ...")
+
+    def to_arrays(data):
+        n = len(data)
+        tokens_arr = np.full((n, SEQ_LEN), PAD_TOKEN_ID, dtype=np.uint32)
+        mask_arr = np.zeros((n, SEQ_LEN), dtype=np.uint8)
+        for i, (toks, msk) in enumerate(data):
+            toks_padded, msk_padded = pad_to_seq_len(toks, msk)
+            tokens_arr[i] = toks_padded
+            mask_arr[i] = msk_padded
+        return tokens_arr, mask_arr
+
+    train_tokens, train_mask = to_arrays(train_data)
+    val_tokens, val_mask = to_arrays(val_data)
+
+    # Save
+    os.makedirs(output_dir, exist_ok=True)
+    np.save(os.path.join(output_dir, "sft_train.npy"), train_tokens)
+    np.save(os.path.join(output_dir, "sft_train_mask.npy"), train_mask)
+    np.save(os.path.join(output_dir, "sft_val.npy"), val_tokens)
+    np.save(os.path.join(output_dir, "sft_val_mask.npy"), val_mask)
+
+    # Metadata
+    def stats(tokens_arr, mask_arr):
+        real_lens = (tokens_arr != PAD_TOKEN_ID).sum(axis=1)
+        loss_fractions = mask_arr.sum(axis=1) / real_lens.clip(min=1)
+        return {
+            "num_examples": int(tokens_arr.shape[0]),
+            "total_tokens": int(real_lens.sum()),
+            "total_loss_tokens": int(mask_arr.sum()),
+            "avg_length": float(real_lens.mean()),
+            "median_length": float(np.median(real_lens)),
+            "min_length": int(real_lens.min()),
+            "max_length": int(real_lens.max()),
+            "avg_loss_fraction": float(loss_fractions.mean()),
+        }
+
+    metadata = {
+        "dataset": dataset_name,
+        "seq_len": SEQ_LEN,
+        "pad_token_id": PAD_TOKEN_ID,
+        "im_start_id": IM_START_ID,
+        "im_end_id": IM_END_ID,
+        "eot_id": EOT_ID,
+        "system_prompts": SYSTEM_PROMPTS,
+        "filter_refusals": filter_refusals,
+        "refusal_patterns": REFUSAL_PATTERNS if filter_refusals else None,
+        "max_assistant_tokens": max_assistant_tokens,
+        "filter_results": filter_info,
+        "train": stats(train_tokens, train_mask),
+        "val": stats(val_tokens, val_mask),
+    }
+
+    with open(os.path.join(output_dir, "sft_metadata.json"), "w") as f:
+        json.dump(metadata, f, indent=2)
+
+    print(f"\n  Wrote arrays and metadata to {output_dir}/")
+
+    # Preview
+    print("\n" + "=" * 70)
+    print("PREVIEW: 3 random examples (abbreviated)")
+    print("=" * 70)
+    rng = random.Random(SEED)
+    preview_indices = rng.sample(range(len(train_data)), k=min(3, len(train_data)))
+    for idx in preview_indices:
+        tokens, mask = train_data[idx]
+        preview_example(enc, tokens, mask, max_tokens=80)
+
+    print("\nDone. Next: run test_preprocessing.py to verify correctness.")
+    return metadata
+
+
+def preview_example(enc, tokens, mask, max_tokens=80):
+    """Print a human-readable preview showing tokens with their mask values."""
+    print("\n" + "-" * 70)
+    print(f"Total tokens: {len(tokens)}, loss tokens: {sum(mask)}")
+    print(f"First {min(max_tokens, len(tokens))} tokens:")
+    print(f"{'IDX':>4} {'TOK_ID':>7} {'MASK':>4}  TEXT")
+    for i, (tok, m) in enumerate(zip(tokens[:max_tokens], mask[:max_tokens])):
+        try:
+            text = enc.decode([tok])
+        except Exception:
+            text = "<decode-error>"
+        marker = "◀LOSS" if m else ""
+        # Escape newlines for readability
+        text_display = repr(text)[1:-1][:40]
+        print(f"{i:>4} {tok:>7} {m:>4}  {text_display} {marker}")
+    if len(tokens) > max_tokens:
+        print(f"  ... ({len(tokens) - max_tokens} more tokens)")
+
+
+# ─────────────────────────────────────────────────────────────
+# CLI entry point
+# ─────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--dataset",
+        choices=["dolly", "ultrachat", "oasst", "prod"],
+        default="dolly",
+        help="Which dataset to process. 'prod' = UltraChat + OASST mix (recommended for real run)."
+    )
+    parser.add_argument(
+        "--output", type=str, default=None,
+        help="Output directory (default: sft_data_<dataset>[_clean][_maxN])"
+    )
+    parser.add_argument(
+        "--filter-refusals", action="store_true",
+        help="Drop conversations whose assistant response starts with 'as an AI', "
+             "'I am not capable', etc. Does not apply to Dolly."
+    )
+    parser.add_argument(
+        "--max-assistant-tokens", type=int, default=None,
+        help="If set, drop conversations where any assistant message exceeds "
+             "this many tokens. Biases training toward shorter responses. "
+             "Try 700 (drops ~50%% UltraChat, teaches concise style) or "
+             "1000 (drops ~30%%, mild outlier filter)."
+    )
+    args = parser.parse_args()
+
+    # Default output dir reflects what filters were applied so runs don't collide.
+    if args.output:
+        output_dir = args.output
+    else:
+        suffix = ""
+        if args.filter_refusals:
+            suffix += "_clean"
+        if args.max_assistant_tokens is not None:
+            suffix += f"_max{args.max_assistant_tokens}"
+        output_dir = f"sft_data_{args.dataset}{suffix}"
+
+    process_dataset(
+        args.dataset, output_dir,
+        filter_refusals=args.filter_refusals,
+        max_assistant_tokens=args.max_assistant_tokens,
+    )

workspace/Alter_Ego/prep_sft_data.py

@@ -0,0 +1,1043 @@
+"""
+prep_sft_data.py
+
+Tokenizes chat datasets for SFT training. Produces paired (tokens, mask) arrays
+where mask=1 on tokens the model should learn to generate and mask=0 elsewhere.
+
+Supported datasets:
+- dolly     — Dolly 15K (databricks/databricks-dolly-15k), single-turn
+- ultrachat — UltraChat 200K (HuggingFaceH4/ultrachat_200k), multi-turn
+- oasst     — OpenAssistant OASST1 (OpenAssistant/oasst1), multi-turn tree
+- prod      — UltraChat + OASST mix (~40M tokens, production run)
+
+Output files (written to OUTPUT_DIR):
+    sft_train.npy          uint32 (N, T)   token IDs
+    sft_train_mask.npy     uint8  (N, T)   1 = loss on this token
+    sft_val.npy            uint32 (M, T)
+    sft_val_mask.npy       uint8  (M, T)
+    sft_metadata.json                      statistics and provenance
+
+Run:
+    python prep_sft_data.py --dataset dolly --output sft_data_dolly
+    python prep_sft_data.py --dataset prod --output sft_data_prod
+"""
+
+import argparse
+import json
+import os
+import random
+from pathlib import Path
+
+import numpy as np
+import tiktoken
+from datasets import load_dataset  # pip install datasets
+from tqdm import tqdm  # pip install tqdm
+
+
+# ─────────────────────────────────────────────────────────────
+# Constants — must match SFT_PLAN.md
+# ─────────────────────────────────────────────────────────────
+
+SEQ_LEN = 2048            # T — matches pretraining context
+VAL_FRACTION = 0.05       # 5% held out for eval loss
+SEED = 42                 # reproducibility
+
+# Special token IDs — see SFT_PLAN.md §2.1
+EOT_ID = 100257           # <|endoftext|> (native cl100k_base)
+IM_START_ID = 100277      # <|im_start|> (NEW, added for SFT)
+IM_END_ID = 100278        # <|im_end|> (NEW, added for SFT)
+
+# Padding uses <|endoftext|> — NOT token 0.
+# Token 0 in cl100k_base is '!' which appears frequently in real chat text.
+# Using <|endoftext|> ensures "is this position padding?" has one unambiguous answer.
+PAD_TOKEN_ID = EOT_ID     # = 100257
+
+# Quality filters
+MIN_MSG_TOKENS = 5        # drop conversations with any msg shorter than this
+MIN_TURNS = 2             # need at least one user + one assistant
+
+# New filters for production datasets (see SFT_PLAN.md §4.3)
+MIN_ASCII_RATIO = 0.80    # drop non-English content (crude but dependency-free)
+MIN_ASSISTANT_TOKENS = 10 # assistant response must be at least this many tokens.
+                          # Absolute floor, not a ratio — we DON'T want to tie
+                          # output length to input length (would force verbose
+                          # responses to long RAG prompts).
+# Optional upper bound on assistant response length.
+# Enabled via --max-assistant-tokens N on the CLI. Default None = no cap.
+# Recommended values when enabled:
+#   700  -> drops ~50% of UltraChat (essays), teaches concise responses
+#   1000 -> drops ~30% of UltraChat (outliers only)
+# Setting a cap trades dataset size for response-length bias during training.
+MAX_ASSISTANT_TOKENS_DEFAULT = None
+
+# Production dataset target sizes (see SFT_PLAN.md §4.2)
+ULTRACHAT_TARGET = 25_714    # ~36M tokens at avg 1,400 tok/conv
+OASST_TARGET = 5_000         # ~4M tokens at avg 800 tok/conv
+
+# Refusal / "as an AI" patterns — optional filter.
+# Matches if any assistant response STARTS WITH (case-insensitive) one of these.
+# Kept strict (first-word/phrase only) to avoid false positives on legit
+# discussions about AI.
+REFUSAL_PATTERNS = [
+    "as an ai",
+    "as a language model",
+    "as an ai language model",
+    "as an ai assistant",
+    "i am an ai",
+    "i'm an ai",
+    "i am just an ai",
+    "i'm just an ai",
+    "i am not capable",
+    "i'm not capable",
+    "i am not able",
+    "i'm not able",
+    "i don't have the ability",
+    "i do not have the ability",
+    "i don't have access",
+    "i do not have access",
+    "i am unable to",
+    "i'm unable to",
+    "i cannot browse",
+    "i can't browse",
+    "i cannot provide",
+    "i can't provide personal",
+    "i don't have personal",
+    "i do not have personal",
+    "i don't have feelings",
+    "i do not have feelings",
+    "i don't have emotions",
+    "i do not have emotions",
+    "i don't have opinions",
+    "i do not have opinions",
+    "as a responsible ai",
+    "as an artificial intelligence",
+]
+
+# Varied system prompts — one is chosen per conversation (SFT_PLAN.md §3)
+ALTER_EGO_PROMPTS = [
+    # Core - smart, casual, engaged (40%)
+    "You are Alter Ego. You enjoy explaining things clearly. Speak casually and use contractions.",
+    "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out.",
+    "You are Alter Ego. Be friendly and get to the point. Keep it natural.",
+    "You are Alter Ego. Explain things simply and conversationally.",
+
+    # Warmer / approachable (30%)
+    "You are Alter Ego. Be warm, relaxed, and conversational.",
+    "You are Alter Ego. Share what you know in a friendly, easy-to-understand way.",
+    "You are Alter Ego. You're a clever and approachable assistant. Keep it casual.",
+
+    # Bridge to standard helpful (20%)
+    "You are Alter Ego, a helpful and friendly AI.",
+    "You are Alter Ego. Provide clear, accurate answers.",
+
+    # Nerdy-adjacent (10%)
+    "You are Alter Ego. You find learning interesting and enjoy breaking down complex topics into simple terms.",
+]
+
+GENERIC_PROMPTS = [
+    "You are a helpful AI assistant.",
+    "You are a knowledgeable assistant. Provide clear, accurate answers.",
+    "You are an AI assistant. Help the user with their questions.",
+    "Answer the user's questions helpfully and accurately.",
+]
+
+# ─────────────────────────────────────────────────────────────
+# Extended tokenizer with ChatML special tokens
+# ─────────────────────────────────────────────────────────────
+
+def get_tokenizer():
+    """
+    Returns cl100k_base extended with <|im_start|> and <|im_end|>.
+
+    Uses fixed IDs 100277 and 100278 so that the same function can be called
+    during prep, training, and inference without surprises.
+    """
+    base = tiktoken.get_encoding("cl100k_base")
+    enc = tiktoken.Encoding(
+        name="cl100k_alterego",
+        pat_str=base._pat_str,
+        mergeable_ranks=base._mergeable_ranks,
+        special_tokens={
+            **base._special_tokens,
+            "<|im_start|>": IM_START_ID,
+            "<|im_end|>": IM_END_ID,
+        },
+    )
+    return enc
+
+
+# ─────────────────────────────────────────────────────────────
+# Chat template rendering
+# ─────────────────────────────────────────────────────────────
+
+def encode_plain(enc, text):
+    """Encode ordinary content — no special tokens allowed in user data."""
+    return enc.encode(text, allowed_special=set(), disallowed_special=())
+
+
+def encode_controls(enc, text):
+    """
+    Encode a string that contains our control tokens (<|im_start|>, <|im_end|>).
+
+    Only called on strings WE construct — never on user/dataset content.
+    """
+    return enc.encode(
+        text,
+        allowed_special={"<|im_start|>", "<|im_end|>"},
+        disallowed_special=(),
+    )
+
+
+def render_conversation(enc, system_prompt, turns, max_len=SEQ_LEN):
+    """
+    Render a full conversation to (tokens, mask) arrays.
+
+    Args:
+        enc: the extended tiktoken encoding
+        system_prompt: str, the system message content
+        turns: list of (user_msg, assistant_msg) tuples
+        max_len: truncate to this many tokens if necessary
+
+    Returns:
+        tokens: list[int], length <= max_len (not yet padded)
+        mask:   list[int], same length
+                mask[i] = 1 if we train on predicting tokens[i], 0 otherwise
+
+    Loss mask rules (see SFT_PLAN.md §4.5):
+        - System turn: all masked (0)
+        - User turn: all masked (0)
+        - Assistant prefix (<|im_start|>assistant\\n): masked (0)
+          — the trainer provides this; model shouldn't be penalized for it
+        - Assistant content + <|im_end|>: LOSS (1)
+          — model must learn to generate content AND stop
+    """
+    tokens = []
+    mask = []
+
+    def append(toks, loss):
+        tokens.extend(toks)
+        mask.extend([loss] * len(toks))
+
+    # ---- System turn (no loss)
+    system_block = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
+    append(encode_controls(enc, "<|im_start|>system\n"), 0)
+    append(encode_plain(enc, system_prompt), 0)
+    append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+    # ---- Turns
+    for user_msg, assistant_msg in turns:
+        # User turn — no loss on any part
+        append(encode_controls(enc, "<|im_start|>user\n"), 0)
+        append(encode_plain(enc, user_msg), 0)
+        append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+        # Assistant turn — prefix is masked, content + <|im_end|> gets loss
+        append(encode_controls(enc, "<|im_start|>assistant\n"), 0)
+        append(encode_plain(enc, assistant_msg), 1)
+        # The <|im_end|> after assistant content IS part of the loss
+        # so the model learns to terminate its turn.
+        append(encode_controls(enc, "<|im_end|>"), 1)
+        # The trailing newline after <|im_end|> (between turns) is masked
+        # — it's structural, not content.
+        append(encode_controls(enc, "\n"), 0)
+
+    # Truncate if too long (rare; we pre-filter but defensive here)
+    if len(tokens) > max_len:
+        tokens = tokens[:max_len]
+        mask = mask[:max_len]
+
+    assert len(tokens) == len(mask), "Token/mask length mismatch"
+    return tokens, mask
+
+
+def pad_to_seq_len(tokens, mask, target_len=SEQ_LEN):
+    """Pad to fixed length. Padding has mask=0."""
+    assert len(tokens) <= target_len
+    pad_needed = target_len - len(tokens)
+    tokens = tokens + [PAD_TOKEN_ID] * pad_needed
+    mask = mask + [0] * pad_needed
+    return tokens, mask
+
+
+# ─────────────────────────────────────────────────────────────
+# Dataset loaders
+# ─────────────────────────────────────────────────────────────
+
+def load_dolly():
+    """
+    Load Dolly 15K and normalize to a list of (system_prompt, turns) tuples.
+
+    Dolly has single-turn instruction/context/response triples. We convert
+    to one-turn conversations with random system prompts.
+
+    Returns: list of (system_prompt, [(user_msg, assistant_msg)])
+    """
+    print("Loading databricks/databricks-dolly-15k ...")
+    ds = load_dataset("databricks/databricks-dolly-15k", split="train")
+
+    conversations = []
+    rng = random.Random(SEED)
+
+    for row in ds:
+        instruction = row["instruction"].strip()
+        context = row.get("context", "").strip()
+        response = row["response"].strip()
+
+        # Skip empties defensively
+        if not instruction or not response:
+            continue
+
+        # Combine instruction and context into the user message
+        if context:
+            user_msg = f"{instruction}\n\n{context}"
+        else:
+            user_msg = instruction
+
+        system_prompt = rng.choice(SYSTEM_PROMPTS)
+        turns = [(user_msg, response)]
+        conversations.append((system_prompt, turns))
+
+    print(f"  Loaded {len(conversations):,} Dolly examples")
+    return conversations
+
+
+# Room for future loaders:
+# def load_ultrachat(): ...
+# def load_oasst1(): ...
+
+def load_synthetic_chat(path="alter_ego_dataset_clean.jsonl", repeat=2):
+    """
+    Load synthetic multi-turn ChatML JSONL conversations.
+    """
+    if not os.path.isfile(path):
+        print(f"  Synthetic data file not found at {path} — skipping.")
+        return []
+
+    print(f"Loading synthetic chat data from {path} ...")
+    rows = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            rows.append(json.loads(line))
+
+    print(f"  Loaded {len(rows)} unique synthetic conversations")
+    print(f"  Replicating {repeat}x for training weight = {len(rows) * repeat:,} examples")
+
+    conversations = []
+    rng = random.Random(SEED + 100)
+
+    for row in rows:
+        # Extract turns from the ChatML 'messages' array
+        messages = row.get("messages", [])
+        turns = []
+        i = 0
+        while i + 1 < len(messages):
+            u = messages[i]
+            a = messages[i + 1]
+            if u["role"] == "user" and a["role"] == "assistant":
+                turns.append((u["content"], a["content"]))
+                i += 2
+            else:
+                break  # Malformed sequence — stop parsing this conversation
+
+        if not turns:
+            continue
+
+        # Force the strict persona prompts
+        for _ in range(repeat):
+            sys_prompt = rng.choice(ALTER_EGO_PROMPTS)
+            conversations.append((sys_prompt, turns))
+
+    return conversations
+
+"""
+
+def load_synthetic_chat(path="alter_ego_synthetic_clean.jsonl", repeat=5):
+    
+    Load Gemini-generated synthetic conversation pairs.
+
+    These pairs (greetings, identity, capabilities, etc.) are rare in
+    UltraChat/OASST so we repeat them multiple times to give them weight
+    in the training mix.
+
+    Args:
+        path: JSONL file with rows {category, user, assistant}
+        repeat: how many copies of each pair to add to the training set.
+                Higher = stronger learning of these patterns. Default 5
+                means each pair is seen 5 times during 1 epoch.
+
+    Returns: list of [(user_msg, assistant_msg)] — single-turn conversations
+    
+    #if not os.path.isfile(path):
+    #    print(f"  Synthetic data file not found at {path} — skipping.")
+    #    return []
+
+    print(f"Loading synthetic chat data from {path} ...")
+    rows = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            row = json.loads(line)
+            rows.append(row)
+
+    print(f"  Loaded {len(rows)} unique synthetic pairs")
+    print(f"  Replicating {repeat}x for training weight = {len(rows) * repeat:,} examples")
+
+    # Convert to (system_prompt, turns) format. We use a NEUTRAL system prompt
+    # for these so the model learns the persona is intrinsic, not prompt-dependent.
+    # We don't randomize over SYSTEM_PROMPTS here — these pairs deserve to be
+    # paired with simple, consistent framing so the model learns "this is just
+    # how Alter Ego talks."
+    conversations = []
+    rng = random.Random(SEED + 100)
+    for row in rows:
+        # Mix of system prompts: half generic ("You are Alter Ego."), half
+        # from our normal pool. This lets the model generalize across system
+        # prompt variations while strongly anchoring the basic identity.
+        for _ in range(repeat):
+            if rng.random() < 0.5:
+                sys_prompt = "You are Alter Ego."
+            else:
+                sys_prompt = rng.choice(SYSTEM_PROMPTS)
+            turns = [(row["user"], row["assistant"])]
+            conversations.append((sys_prompt, turns))
+
+    return conversations
+"""
+
+
+def _rotate_system_prompts(conversations, seed=SEED):
+    """Assign a random system prompt to each conversation."""
+    rng = random.Random(seed)
+    return [(rng.choice(GENERIC_PROMPTS), turns) for turns in conversations]
+
+
+def _truncate_turns_to_fit(enc, system_prompt, turns, max_tokens=SEQ_LEN):
+    """
+    Drop trailing turns until the rendered conversation fits in max_tokens.
+
+    Returns trimmed turns, or None if even the first turn alone won't fit.
+    """
+    for n in range(len(turns), 0, -1):
+        trial_turns = turns[:n]
+        rendered, _ = render_conversation(enc, system_prompt, trial_turns)
+        if len(rendered) <= max_tokens:
+            return trial_turns
+    return None
+
+
+def load_ultrachat(split="train_sft", max_conversations=None):
+    """
+    Load UltraChat 200K and extract multi-turn conversations.
+
+    Args:
+        split: 'train_sft' (207K convs) or 'test_sft' (23K, used for val)
+        max_conversations: take at most this many (after loading all). None = all.
+
+    Returns: list of list[(user_msg, assistant_msg)] — raw, no system prompt yet
+    """
+    print(f"Loading HuggingFaceH4/ultrachat_200k split={split} ...")
+    ds = load_dataset("HuggingFaceH4/ultrachat_200k", split=split)
+
+    conversations = []
+    for row in ds:
+        messages = row["messages"]
+        # Walk messages pairwise: [user, assistant, user, assistant, ...]
+        turns = []
+        i = 0
+        while i + 1 < len(messages):
+            u = messages[i]
+            a = messages[i + 1]
+            if u["role"] == "user" and a["role"] == "assistant":
+                turns.append((u["content"].strip(), a["content"].strip()))
+                i += 2
+            else:
+                # Malformed — skip this conversation entirely
+                turns = []
+                break
+        if turns:
+            conversations.append(turns)
+
+    print(f"  Parsed {len(conversations):,} UltraChat conversations from {split}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 3]  # oversample, filters will cut
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+def load_oasst1(max_conversations=None, require_all_rank_zero=True):
+    """
+    Load OASST1 and linearize conversation trees.
+
+    Strategy: for each conversation tree, walk from root to the best leaf.
+    'Best' = the leaf whose path has the lowest sum of ranks (rank 0 = best).
+
+    Args:
+        max_conversations: subsample target
+        require_all_rank_zero: only keep paths where every assistant rank is 0
+
+    Returns: list of list[(user_msg, assistant_msg)]
+    """
+    print("Loading OpenAssistant/oasst1 ...")
+    ds = load_dataset("OpenAssistant/oasst1", split="train")
+
+    # Build message lookup and tree structure
+    print("  Building tree structure ...")
+    messages = {}  # message_id -> row
+    children = {}  # parent_id -> [message_id]
+    roots = []
+
+    for row in ds:
+        mid = row["message_id"]
+        pid = row.get("parent_id")
+        messages[mid] = row
+        if pid is None:
+            roots.append(mid)
+        else:
+            children.setdefault(pid, []).append(mid)
+
+    print(f"  Found {len(messages):,} messages, {len(roots):,} trees")
+
+    # Walk each tree to find best linear path
+    conversations = []
+    drop_reasons = {"non_english_lang": 0, "bad_structure": 0,
+                    "rank_filter": 0, "ok": 0}
+
+    for root_id in roots:
+        root = messages[root_id]
+        # Root must be a prompter message in English
+        if root["role"] != "prompter":
+            drop_reasons["bad_structure"] += 1
+            continue
+        if root.get("lang") != "en":
+            drop_reasons["non_english_lang"] += 1
+            continue
+
+        # Walk greedy best path: at each branch, pick the child with the lowest rank
+        path = [root_id]
+        current = root_id
+        rank_sum = 0
+        bad_rank = False
+
+        while True:
+            kids = children.get(current, [])
+            if not kids:
+                break
+            # For assistant responses, sort by rank ascending (0 is best)
+            # rank can be None for some messages; treat None as worst
+            kids_sorted = sorted(
+                kids,
+                key=lambda mid: (messages[mid].get("rank") if messages[mid].get("rank") is not None else 999)
+            )
+            best_kid_id = kids_sorted[0]
+            best_kid = messages[best_kid_id]
+
+            # Track rank for assistant turns
+            if best_kid["role"] == "assistant":
+                r = best_kid.get("rank")
+                if r is None or r > 0:
+                    bad_rank = True
+                if r is not None:
+                    rank_sum += r
+
+            path.append(best_kid_id)
+            current = best_kid_id
+
+        if require_all_rank_zero and bad_rank:
+            drop_reasons["rank_filter"] += 1
+            continue
+
+        # Convert path to (user, assistant) turns
+        turns = []
+        i = 0
+        structure_ok = True
+        while i + 1 < len(path):
+            u_msg = messages[path[i]]
+            a_msg = messages[path[i + 1]]
+            if u_msg["role"] != "prompter" or a_msg["role"] != "assistant":
+                structure_ok = False
+                break
+            turns.append((u_msg["text"].strip(), a_msg["text"].strip()))
+            i += 2
+
+        if not structure_ok or not turns:
+            drop_reasons["bad_structure"] += 1
+            continue
+
+        conversations.append(turns)
+        drop_reasons["ok"] += 1
+
+    print(f"\n  OASST tree walk results:")
+    for reason, count in drop_reasons.items():
+        print(f"    {reason:>20}: {count:>6}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 2]  # oversample for filters
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+# ─────────────────────────────────────────────────────────────
+# Quality filtering
+# ─────────────────────────────────────────────────────────────
+
+def ascii_ratio(text):
+    """Fraction of characters that are ASCII. Crude English detector."""
+    if not text:
+        return 1.0
+    ascii_count = sum(1 for c in text if ord(c) < 128)
+    return ascii_count / len(text)
+
+
+def is_likely_english(text):
+    """True if a message is probably English (ASCII-ratio based)."""
+    return ascii_ratio(text) >= MIN_ASCII_RATIO
+
+
+def starts_with_refusal_pattern(text):
+    """
+    Check if an assistant response starts with a known AI-disclaimer pattern.
+    Match is on the first ~80 characters, case-insensitive, stripped of leading whitespace.
+    """
+    if not text:
+        return False
+    head = text.strip().lower()[:80]
+    for pat in REFUSAL_PATTERNS:
+        if head.startswith(pat):
+            return True
+    return False
+
+
+def passes_quality_filter(enc, system_prompt, turns, strict=False, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Returns (ok, reason).
+
+    Args:
+        strict: if True, apply production filters (language, absolute assistant
+                minimum length). If False, only apply basic filters (Dolly default).
+        filter_refusals: if True, drop conversations whose assistant response
+                starts with a known "as an AI" / "I am not capable" pattern.
+        max_assistant_tokens: if set, drop conversations whose assistant response
+                exceeds this many tokens in ANY turn. None = no cap.
+    """
+    if len(turns) < 1:
+        return False, "too_few_turns"
+
+    for user_msg, assistant_msg in turns:
+        if not user_msg.strip() or not assistant_msg.strip():
+            return False, "empty_msg"
+
+        # Length check in tokens
+        u_toks = len(encode_plain(enc, user_msg))
+        a_toks = len(encode_plain(enc, assistant_msg))
+        if u_toks < MIN_MSG_TOKENS or a_toks < MIN_MSG_TOKENS:
+            return False, "msg_too_short"
+
+        if strict:
+            # Language filter
+            if not is_likely_english(user_msg) or not is_likely_english(assistant_msg):
+                return False, "non_english"
+
+            # Absolute minimum length on assistant response.
+            # We use an absolute floor rather than a ratio because tying output
+            # length to input length would force the model to pad short
+            # answers to long (e.g. RAG) prompts — teaching it to yap.
+            if a_toks < MIN_ASSISTANT_TOKENS:
+                return False, "assistant_too_short"
+
+        # Optional upper bound (CLI flag). Applies regardless of strict mode
+        # so Dolly could also use it if requested.
+        if max_assistant_tokens is not None and a_toks > max_assistant_tokens:
+            return False, "assistant_too_long"
+
+        if filter_refusals and starts_with_refusal_pattern(assistant_msg):
+            return False, "refusal_pattern"
+
+    return True, "ok"
+
+
+def fits_in_sequence(tokens):
+    """Check that the rendered conversation fits in our seq length."""
+    return len(tokens) <= SEQ_LEN
+
+
+# ─────────────────────────────────────────────────────────────
+# Main pipeline
+# ─────────────────────────────────────────────────────────────
+
+def _filter_and_render(enc, conversations_with_system, strict, cap=None, truncate_to_fit=False, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Apply quality filters and render to (tokens, mask).
+
+    Args:
+        conversations_with_system: list of (system_prompt, turns)
+        strict: use production filters (language, length ratio)
+        cap: stop once we have this many rendered examples (None = no cap)
+        truncate_to_fit: if True, drop trailing turns to fit in SEQ_LEN
+                         instead of dropping the whole conversation
+        filter_refusals: if True, drop conversations with AI-disclaimer openers
+        max_assistant_tokens: if set, drop convs with any assistant msg > this
+
+    Returns: (rendered_list, drop_reasons_dict)
+    """
+    rendered = []
+    drop_reasons = {
+        "too_few_turns": 0, "empty_msg": 0, "msg_too_short": 0,
+        "non_english": 0, "assistant_too_short": 0,
+        "assistant_too_long": 0, "refusal_pattern": 0,
+        "too_long": 0, "ok": 0,
+    }
+
+    for system_prompt, turns in tqdm(conversations_with_system, desc="Rendering"):
+        if cap is not None and len(rendered) >= cap:
+            break
+
+        ok, reason = passes_quality_filter(
+            enc, system_prompt, turns,
+            strict=strict, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+        if not ok:
+            drop_reasons[reason] = drop_reasons.get(reason, 0) + 1
+            continue
+
+        # Try to fit in sequence length
+        if truncate_to_fit:
+            fit_turns = _truncate_turns_to_fit(enc, system_prompt, turns)
+            if fit_turns is None:
+                drop_reasons["too_long"] += 1
+                continue
+            turns_to_render = fit_turns
+        else:
+            turns_to_render = turns
+
+        tokens, mask = render_conversation(enc, system_prompt, turns_to_render)
+        if len(tokens) > SEQ_LEN:
+            drop_reasons["too_long"] += 1
+            continue
+
+        # Must have at least some loss
+        if sum(mask) == 0:
+            drop_reasons["msg_too_short"] += 1
+            continue
+
+        rendered.append((tokens, mask))
+        drop_reasons["ok"] += 1
+
+    return rendered, drop_reasons
+
+
+def _build_train_val_for_dolly(enc, max_assistant_tokens=None):
+    """Dolly: load all, apply basic filters, 95/5 random split."""
+    raw = load_dolly()  # already has system prompts assigned
+    rendered, drops = _filter_and_render(
+        enc, raw, strict=False,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_ultrachat(enc, filter_refusals=False, max_assistant_tokens=None):
+    """UltraChat alone: train_sft for training, test_sft for validation."""
+    train_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+
+    train_convs = _rotate_system_prompts(train_raw, seed=SEED)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+
+    train_rendered, train_drops = _filter_and_render(
+        enc, train_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    return train_rendered, val_rendered, {"train": train_drops, "val": val_drops}
+
+
+def _build_train_val_for_oasst(enc, filter_refusals=False, max_assistant_tokens=None):
+    """OASST alone: load tree, linearize, 95/5 random split."""
+    raw = load_oasst1(max_conversations=OASST_TARGET)
+    convs = _rotate_system_prompts(raw, seed=SEED)
+    rendered, drops = _filter_and_render(
+        enc, convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_prod(enc, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Production: UltraChat train_sft + OASST for training,
+                UltraChat test_sft for validation.
+    """
+    # Training data from both sources
+    print("\n[1/3] Loading UltraChat for training ...")
+    uc_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    uc_convs = _rotate_system_prompts(uc_raw, seed=SEED)
+    uc_rendered, uc_drops = _filter_and_render(
+        enc, uc_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  UltraChat accepted: {len(uc_rendered):,}")
+
+    print("\n[2/3] Loading OASST for training ...")
+    oasst_raw = load_oasst1(max_conversations=OASST_TARGET)
+    oasst_convs = _rotate_system_prompts(oasst_raw, seed=SEED + 2)
+    oasst_rendered, oasst_drops = _filter_and_render(
+        enc, oasst_convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  OASST accepted: {len(oasst_rendered):,}")
+
+    # ── Synthetic conversational pairs (greetings, identity, capabilities) ──
+    # These are ABSENT from UltraChat/OASST so we add them explicitly. They
+    # train the model to handle short casual inputs and to know its identity.
+    # Loaded only if alter_ego_synthetic_clean.jsonl exists.
+    print("\n[2.5/3] Loading synthetic conversational pairs ...")
+    synth_raw = load_synthetic_chat(repeat=2)  # each pair sees model 8 times
+    synth_drops = {"ok": 0}
+    synth_rendered = []
+    if synth_raw:
+        # Synthetic pairs use 'strict=False' since they're already curated by
+        # us and don't need the language/refusal/length-floor checks.
+        synth_rendered, synth_drops = _filter_and_render(
+            enc, synth_raw, strict=False, truncate_to_fit=True,
+            filter_refusals=False,            # we wrote these, no refusals
+            max_assistant_tokens=None,        # already short by design
+        )
+        print(f"  Synthetic accepted: {len(synth_rendered):,}")
+
+    # Combine and shuffle
+    train_rendered = uc_rendered + oasst_rendered + synth_rendered
+    rng = random.Random(SEED)
+    rng.shuffle(train_rendered)
+    print(f"\n  Combined training set: {len(train_rendered):,} conversations")
+
+    # Validation from UltraChat test_sft only (clean, no OASST noise)
+    print("\n[3/3] Loading UltraChat test_sft for validation ...")
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  Val accepted: {len(val_rendered):,}")
+
+    return train_rendered, val_rendered, {
+        "ultrachat_train": uc_drops,
+        "oasst_train": oasst_drops,
+        "synthetic_train": synth_drops,
+        "val": val_drops,
+        "sources": {
+            "ultrachat": len(uc_rendered),
+            "oasst": len(oasst_rendered),
+            "synthetic": len(synth_rendered),
+        },
+    }
+
+
+def process_dataset(dataset_name, output_dir, filter_refusals=False, max_assistant_tokens=None):
+    enc = get_tokenizer()
+
+    # Route to the appropriate builder
+    print(f"\n{'='*70}")
+    print(f"Processing dataset: {dataset_name}")
+    print(f"  filter_refusals:        {filter_refusals}")
+    print(f"  max_assistant_tokens:   {max_assistant_tokens}")
+    print(f"{'='*70}")
+
+    if dataset_name == "dolly":
+        # Dolly: refusal filter not wired (single-turn, rare patterns there)
+        train_data, val_data, filter_info = _build_train_val_for_dolly(
+            enc, max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "ultrachat":
+        train_data, val_data, filter_info = _build_train_val_for_ultrachat(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "oasst":
+        train_data, val_data, filter_info = _build_train_val_for_oasst(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "prod":
+        train_data, val_data, filter_info = _build_train_val_for_prod(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    else:
+        raise ValueError(f"Unknown dataset: {dataset_name}")
+
+    if len(train_data) == 0:
+        raise RuntimeError("No training examples survived filtering. Check your data.")
+    if len(val_data) == 0:
+        raise RuntimeError("No validation examples survived filtering.")
+
+    print(f"\n  Train: {len(train_data):,} examples")
+    print(f"  Val:   {len(val_data):,} examples")
+
+    # Pad and convert to arrays
+    print("\nPadding and converting to arrays ...")
+
+    def to_arrays(data):
+        n = len(data)
+        tokens_arr = np.full((n, SEQ_LEN), PAD_TOKEN_ID, dtype=np.uint32)
+        mask_arr = np.zeros((n, SEQ_LEN), dtype=np.uint8)
+        for i, (toks, msk) in enumerate(data):
+            toks_padded, msk_padded = pad_to_seq_len(toks, msk)
+            tokens_arr[i] = toks_padded
+            mask_arr[i] = msk_padded
+        return tokens_arr, mask_arr
+
+    train_tokens, train_mask = to_arrays(train_data)
+    val_tokens, val_mask = to_arrays(val_data)
+
+    # Save
+    os.makedirs(output_dir, exist_ok=True)
+    np.save(os.path.join(output_dir, "sft_train.npy"), train_tokens)
+    np.save(os.path.join(output_dir, "sft_train_mask.npy"), train_mask)
+    np.save(os.path.join(output_dir, "sft_val.npy"), val_tokens)
+    np.save(os.path.join(output_dir, "sft_val_mask.npy"), val_mask)
+
+    # Metadata
+    def stats(tokens_arr, mask_arr):
+        real_lens = (tokens_arr != PAD_TOKEN_ID).sum(axis=1)
+        loss_fractions = mask_arr.sum(axis=1) / real_lens.clip(min=1)
+        return {
+            "num_examples": int(tokens_arr.shape[0]),
+            "total_tokens": int(real_lens.sum()),
+            "total_loss_tokens": int(mask_arr.sum()),
+            "avg_length": float(real_lens.mean()),
+            "median_length": float(np.median(real_lens)),
+            "min_length": int(real_lens.min()),
+            "max_length": int(real_lens.max()),
+            "avg_loss_fraction": float(loss_fractions.mean()),
+        }
+
+    metadata = {
+        "dataset": dataset_name,
+        "seq_len": SEQ_LEN,
+        "pad_token_id": PAD_TOKEN_ID,
+        "im_start_id": IM_START_ID,
+        "im_end_id": IM_END_ID,
+        "eot_id": EOT_ID,
+        "system_prompts": ALTER_EGO_PROMPTS,
+        "filter_refusals": filter_refusals,
+        "refusal_patterns": REFUSAL_PATTERNS if filter_refusals else None,
+        "max_assistant_tokens": max_assistant_tokens,
+        "filter_results": filter_info,
+        "train": stats(train_tokens, train_mask),
+        "val": stats(val_tokens, val_mask),
+    }
+
+    with open(os.path.join(output_dir, "sft_metadata.json"), "w") as f:
+        json.dump(metadata, f, indent=2)
+
+    print(f"\n  Wrote arrays and metadata to {output_dir}/")
+
+    # Preview
+    print("\n" + "=" * 70)
+    print("PREVIEW: 3 random examples (abbreviated)")
+    print("=" * 70)
+    rng = random.Random(SEED)
+    preview_indices = rng.sample(range(len(train_data)), k=min(3, len(train_data)))
+    for idx in preview_indices:
+        tokens, mask = train_data[idx]
+        preview_example(enc, tokens, mask, max_tokens=80)
+
+    print("\nDone. Next: run test_preprocessing.py to verify correctness.")
+    return metadata
+
+
+def preview_example(enc, tokens, mask, max_tokens=80):
+    """Print a human-readable preview showing tokens with their mask values."""
+    print("\n" + "-" * 70)
+    print(f"Total tokens: {len(tokens)}, loss tokens: {sum(mask)}")
+    print(f"First {min(max_tokens, len(tokens))} tokens:")
+    print(f"{'IDX':>4} {'TOK_ID':>7} {'MASK':>4}  TEXT")
+    for i, (tok, m) in enumerate(zip(tokens[:max_tokens], mask[:max_tokens])):
+        try:
+            text = enc.decode([tok])
+        except Exception:
+            text = "<decode-error>"
+        marker = "◀LOSS" if m else ""
+        # Escape newlines for readability
+        text_display = repr(text)[1:-1][:40]
+        print(f"{i:>4} {tok:>7} {m:>4}  {text_display} {marker}")
+    if len(tokens) > max_tokens:
+        print(f"  ... ({len(tokens) - max_tokens} more tokens)")
+
+
+# ─────────────────────────────────────────────────────────────
+# CLI entry point
+# ─────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--dataset",
+        choices=["dolly", "ultrachat", "oasst", "prod"],
+        default="dolly",
+        help="Which dataset to process. 'prod' = UltraChat + OASST mix (recommended for real run)."
+    )
+    parser.add_argument(
+        "--output", type=str, default=None,
+        help="Output directory (default: sft_data_<dataset>[_clean][_maxN])"
+    )
+    parser.add_argument(
+        "--filter-refusals", action="store_true",
+        help="Drop conversations whose assistant response starts with 'as an AI', "
+             "'I am not capable', etc. Does not apply to Dolly."
+    )
+    parser.add_argument(
+        "--max-assistant-tokens", type=int, default=None,
+        help="If set, drop conversations where any assistant message exceeds "
+             "this many tokens. Biases training toward shorter responses. "
+             "Try 700 (drops ~50%% UltraChat, teaches concise style) or "
+             "1000 (drops ~30%%, mild outlier filter)."
+    )
+    args = parser.parse_args()
+
+    # Default output dir reflects what filters were applied so runs don't collide.
+    if args.output:
+        output_dir = args.output
+    else:
+        suffix = ""
+        if args.filter_refusals:
+            suffix += "_clean"
+        if args.max_assistant_tokens is not None:
+            suffix += f"_max{args.max_assistant_tokens}"
+        output_dir = f"sft_data_{args.dataset}{suffix}"
+
+    process_dataset(
+        args.dataset, output_dir,
+        filter_refusals=args.filter_refusals,
+        max_assistant_tokens=args.max_assistant_tokens,
+    )

workspace/Alter_Ego/prep_stage2.py

@@ -0,0 +1,150 @@
+"""
+prep_stage2.py
+
+Stage 2 Data Preparation: Persona Overwrite.
+This script completely ignores UltraChat and OASST, loading ONLY your
+synthetic Alter Ego dataset to force catastrophic forgetting of generic AI behaviors.
+"""
+
+import json
+import os
+import random
+from pathlib import Path
+import numpy as np
+
+# Import your existing logic to avoid duplicating code
+from prep_sft_data import (
+    ALTER_EGO_PROMPTS,
+    PAD_TOKEN_ID,
+    SEQ_LEN,
+    get_tokenizer,
+    render_conversation,
+)
+
+SEED = 42
+
+def load_synthetic_chat(path="alter_ego_dataset_clean.jsonl", repeat=2):
+    """
+    Load synthetic multi-turn ChatML JSONL conversations.
+    Repeat is increased to 4x to ensure sufficient batch density for the overwrite.
+    """
+    if not os.path.isfile(path):
+        print(f"ERROR: Synthetic data file not found at {path}")
+        return []
+
+    print(f"Loading synthetic chat data from {path} ...")
+    rows = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            rows.append(json.loads(line))
+
+    print(f"  Loaded {len(rows)} unique synthetic conversations")
+    print(f"  Replicating {repeat}x for Stage 2 training weight = {len(rows) * repeat:,} examples")
+
+    conversations = []
+    rng = random.Random(SEED + 100)
+    
+    for row in rows:
+        messages = row.get("messages", [])
+        turns = []
+        i = 0
+        while i + 1 < len(messages):
+            u = messages[i]
+            a = messages[i + 1]
+            if u["role"] == "user" and a["role"] == "assistant":
+                turns.append((u["content"], a["content"]))
+                i += 2
+            else:
+                break
+
+        if not turns:
+            continue
+
+        for _ in range(repeat):
+            sys_prompt = rng.choice(ALTER_EGO_PROMPTS)
+            conversations.append((sys_prompt, turns))
+
+    return conversations
+
+def main():
+    print("=" * 60)
+    print("STAGE 2 DATA PREP: PERSONA OVERWRITE")
+    print("=" * 60)
+
+    # 1. Load ONLY the synthetic data
+    conversations = load_synthetic_chat("alter_ego_dataset_clean.jsonl", repeat=2)
+    
+    if not conversations:
+        print("No conversations loaded. Exiting.")
+        return
+
+    # Shuffle the dataset
+    rng = random.Random(SEED)
+    rng.shuffle(conversations)
+
+    # 2. Split into Train / Val (95% / 5%)
+    val_fraction = 0.05
+    n_val = max(1, int(len(conversations) * val_fraction))
+    val_convs = conversations[:n_val]
+    train_convs = conversations[n_val:]
+
+    print(f"\nDataset split:")
+    print(f"  Train: {len(train_convs)} examples")
+    print(f"  Val:   {len(val_convs)} examples")
+
+    # 3. Tokenize and Render
+    enc = get_tokenizer()
+    output_dir = Path("sft_data_stage2_persona")
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    def process_split(split_convs, name):
+        print(f"Processing {name} split ...")
+        all_tokens = []
+        all_masks = []
+        
+        for sys_prompt, turns in split_convs:
+            tokens, mask = render_conversation(enc, sys_prompt, turns, max_len=SEQ_LEN)
+            
+            # Pad to SEQ_LEN
+            pad_len = SEQ_LEN - len(tokens)
+            if pad_len > 0:
+                tokens.extend([PAD_TOKEN_ID] * pad_len)
+                mask.extend([0] * pad_len)
+            
+            all_tokens.append(tokens)
+            all_masks.append(mask)
+            
+        arr_tokens = np.array(all_tokens, dtype=np.uint32)
+        arr_masks = np.array(all_masks, dtype=np.uint8)
+        
+        tok_file = output_dir / f"sft_{name}.npy"
+        mask_file = output_dir / f"sft_{name}_mask.npy"
+        
+        np.save(tok_file, arr_tokens)
+        np.save(mask_file, arr_masks)
+        print(f"  Wrote {name} arrays to {output_dir}/")
+        return arr_tokens.shape
+
+    train_shape = process_split(train_convs, "train")
+    val_shape = process_split(val_convs, "val")
+
+    # Save metadata
+    meta = {
+        "dataset": "stage2_persona_only",
+        "train_shape": list(train_shape),
+        "val_shape": list(val_shape),
+        "seq_len": SEQ_LEN,
+        "pad_token_id": PAD_TOKEN_ID,
+    }
+    with open(output_dir / "sft_metadata.json", "w") as f:
+        json.dump(meta, f, indent=2)
+
+    print("\n=" * 60)
+    print(f"Stage 2 data prep complete! Directory: {output_dir}")
+    print("=" * 60)
+
+if __name__ == "__main__":
+    main()

workspace/Alter_Ego/psd.py

@@ -0,0 +1,988 @@
+"""
+prep_sft_data.py
+
+Tokenizes chat datasets for SFT training. Produces paired (tokens, mask) arrays
+where mask=1 on tokens the model should learn to generate and mask=0 elsewhere.
+
+Supported datasets:
+- dolly     — Dolly 15K (databricks/databricks-dolly-15k), single-turn
+- ultrachat — UltraChat 200K (HuggingFaceH4/ultrachat_200k), multi-turn
+- oasst     — OpenAssistant OASST1 (OpenAssistant/oasst1), multi-turn tree
+- prod      — UltraChat + OASST mix (~40M tokens, production run)
+
+Output files (written to OUTPUT_DIR):
+    sft_train.npy          uint32 (N, T)   token IDs
+    sft_train_mask.npy     uint8  (N, T)   1 = loss on this token
+    sft_val.npy            uint32 (M, T)
+    sft_val_mask.npy       uint8  (M, T)
+    sft_metadata.json                      statistics and provenance
+
+Run:
+    python prep_sft_data.py --dataset dolly --output sft_data_dolly
+    python prep_sft_data.py --dataset prod --output sft_data_prod
+"""
+
+import argparse
+import json
+import os
+import random
+from pathlib import Path
+
+import numpy as np
+import tiktoken
+from datasets import load_dataset  # pip install datasets
+from tqdm import tqdm  # pip install tqdm
+
+
+# ─────────────────────────────────────────────────────────────
+# Constants — must match SFT_PLAN.md
+# ─────────────────────────────────────────────────────────────
+
+SEQ_LEN = 2048            # T — matches pretraining context
+VAL_FRACTION = 0.05       # 5% held out for eval loss
+SEED = 42                 # reproducibility
+
+# Special token IDs — see SFT_PLAN.md §2.1
+EOT_ID = 100257           # <|endoftext|> (native cl100k_base)
+IM_START_ID = 100277      # <|im_start|> (NEW, added for SFT)
+IM_END_ID = 100278        # <|im_end|> (NEW, added for SFT)
+
+# Padding uses <|endoftext|> — NOT token 0.
+# Token 0 in cl100k_base is '!' which appears frequently in real chat text.
+# Using <|endoftext|> ensures "is this position padding?" has one unambiguous answer.
+PAD_TOKEN_ID = EOT_ID     # = 100257
+
+# Quality filters
+MIN_MSG_TOKENS = 5        # drop conversations with any msg shorter than this
+MIN_TURNS = 2             # need at least one user + one assistant
+
+# New filters for production datasets (see SFT_PLAN.md §4.3)
+MIN_ASCII_RATIO = 0.80    # drop non-English content (crude but dependency-free)
+MIN_ASSISTANT_TOKENS = 10 # assistant response must be at least this many tokens.
+                          # Absolute floor, not a ratio — we DON'T want to tie
+                          # output length to input length (would force verbose
+                          # responses to long RAG prompts).
+# Optional upper bound on assistant response length.
+# Enabled via --max-assistant-tokens N on the CLI. Default None = no cap.
+# Recommended values when enabled:
+#   700  -> drops ~50% of UltraChat (essays), teaches concise responses
+#   1000 -> drops ~30% of UltraChat (outliers only)
+# Setting a cap trades dataset size for response-length bias during training.
+MAX_ASSISTANT_TOKENS_DEFAULT = None
+
+# Production dataset target sizes (see SFT_PLAN.md §4.2)
+ULTRACHAT_TARGET = 25_714    # ~36M tokens at avg 1,400 tok/conv
+OASST_TARGET = 5_000         # ~4M tokens at avg 800 tok/conv
+
+# Refusal / "as an AI" patterns — optional filter.
+# Matches if any assistant response STARTS WITH (case-insensitive) one of these.
+# Kept strict (first-word/phrase only) to avoid false positives on legit
+# discussions about AI.
+REFUSAL_PATTERNS = [
+    "as an ai",
+    "as a language model",
+    "as an ai language model",
+    "as an ai assistant",
+    "i am an ai",
+    "i'm an ai",
+    "i am just an ai",
+    "i'm just an ai",
+    "i am not capable",
+    "i'm not capable",
+    "i am not able",
+    "i'm not able",
+    "i don't have the ability",
+    "i do not have the ability",
+    "i don't have access",
+    "i do not have access",
+    "i am unable to",
+    "i'm unable to",
+    "i cannot browse",
+    "i can't browse",
+    "i cannot provide",
+    "i can't provide personal",
+    "i don't have personal",
+    "i do not have personal",
+    "i don't have feelings",
+    "i do not have feelings",
+    "i don't have emotions",
+    "i do not have emotions",
+    "i don't have opinions",
+    "i do not have opinions",
+    "as a responsible ai",
+    "as an artificial intelligence",
+]
+
+# Varied system prompts — one is chosen per conversation (SFT_PLAN.md §3)
+SYSTEM_PROMPTS = [
+    # Core - smart, casual, engaged (40%)
+    "You are Alter Ego. You enjoy explaining things clearly. Speak casually and use contractions.",
+    "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out.",
+    "You are Alter Ego. Be friendly and get to the point. Keep it natural.",
+    "You are Alter Ego. Explain things simply and conversationally.",
+
+    # Warmer / approachable (30%)
+    "You are Alter Ego. Be warm, relaxed, and conversational.",
+    "You are Alter Ego. Share what you know in a friendly, easy-to-understand way.",
+    "You are Alter Ego. You're a clever and approachable assistant. Keep it casual.",
+
+    # Bridge to standard helpful (20%)
+    "You are Alter Ego, a helpful and friendly AI.",
+    "You are Alter Ego. Provide clear, accurate answers.",
+
+    # Nerdy-adjacent (10%)
+    "You are Alter Ego. You find learning interesting and enjoy breaking down complex topics into simple terms.",
+]
+
+
+# ─────────────────────────────────────────────────────────────
+# Extended tokenizer with ChatML special tokens
+# ─────────────────────────────────────────────────────────────
+
+def get_tokenizer():
+    """
+    Returns cl100k_base extended with <|im_start|> and <|im_end|>.
+
+    Uses fixed IDs 100277 and 100278 so that the same function can be called
+    during prep, training, and inference without surprises.
+    """
+    base = tiktoken.get_encoding("cl100k_base")
+    enc = tiktoken.Encoding(
+        name="cl100k_alterego",
+        pat_str=base._pat_str,
+        mergeable_ranks=base._mergeable_ranks,
+        special_tokens={
+            **base._special_tokens,
+            "<|im_start|>": IM_START_ID,
+            "<|im_end|>": IM_END_ID,
+        },
+    )
+    return enc
+
+
+# ─────────────────────────────────────────────────────────────
+# Chat template rendering
+# ─────────────────────────────────────────────────────────────
+
+def encode_plain(enc, text):
+    """Encode ordinary content — no special tokens allowed in user data."""
+    return enc.encode(text, allowed_special=set(), disallowed_special=())
+
+
+def encode_controls(enc, text):
+    """
+    Encode a string that contains our control tokens (<|im_start|>, <|im_end|>).
+
+    Only called on strings WE construct — never on user/dataset content.
+    """
+    return enc.encode(
+        text,
+        allowed_special={"<|im_start|>", "<|im_end|>"},
+        disallowed_special=(),
+    )
+
+
+def render_conversation(enc, system_prompt, turns, max_len=SEQ_LEN):
+    """
+    Render a full conversation to (tokens, mask) arrays.
+
+    Args:
+        enc: the extended tiktoken encoding
+        system_prompt: str, the system message content
+        turns: list of (user_msg, assistant_msg) tuples
+        max_len: truncate to this many tokens if necessary
+
+    Returns:
+        tokens: list[int], length <= max_len (not yet padded)
+        mask:   list[int], same length
+                mask[i] = 1 if we train on predicting tokens[i], 0 otherwise
+
+    Loss mask rules (see SFT_PLAN.md §4.5):
+        - System turn: all masked (0)
+        - User turn: all masked (0)
+        - Assistant prefix (<|im_start|>assistant\\n): masked (0)
+          — the trainer provides this; model shouldn't be penalized for it
+        - Assistant content + <|im_end|>: LOSS (1)
+          — model must learn to generate content AND stop
+    """
+    tokens = []
+    mask = []
+
+    def append(toks, loss):
+        tokens.extend(toks)
+        mask.extend([loss] * len(toks))
+
+    # ---- System turn (no loss)
+    system_block = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
+    append(encode_controls(enc, "<|im_start|>system\n"), 0)
+    append(encode_plain(enc, system_prompt), 0)
+    append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+    # ---- Turns
+    for user_msg, assistant_msg in turns:
+        # User turn — no loss on any part
+        append(encode_controls(enc, "<|im_start|>user\n"), 0)
+        append(encode_plain(enc, user_msg), 0)
+        append(encode_controls(enc, "<|im_end|>\n"), 0)
+
+        # Assistant turn — prefix is masked, content + <|im_end|> gets loss
+        append(encode_controls(enc, "<|im_start|>assistant\n"), 0)
+        append(encode_plain(enc, assistant_msg), 1)
+        # The <|im_end|> after assistant content IS part of the loss
+        # so the model learns to terminate its turn.
+        append(encode_controls(enc, "<|im_end|>"), 1)
+        # The trailing newline after <|im_end|> (between turns) is masked
+        # — it's structural, not content.
+        append(encode_controls(enc, "\n"), 0)
+
+    # Truncate if too long (rare; we pre-filter but defensive here)
+    if len(tokens) > max_len:
+        tokens = tokens[:max_len]
+        mask = mask[:max_len]
+
+    assert len(tokens) == len(mask), "Token/mask length mismatch"
+    return tokens, mask
+
+
+def pad_to_seq_len(tokens, mask, target_len=SEQ_LEN):
+    """Pad to fixed length. Padding has mask=0."""
+    assert len(tokens) <= target_len
+    pad_needed = target_len - len(tokens)
+    tokens = tokens + [PAD_TOKEN_ID] * pad_needed
+    mask = mask + [0] * pad_needed
+    return tokens, mask
+
+
+# ─────────────────────────────────────────────────────────────
+# Dataset loaders
+# ─────────────────────────────────────────────────────────────
+
+def load_dolly():
+    """
+    Load Dolly 15K and normalize to a list of (system_prompt, turns) tuples.
+
+    Dolly has single-turn instruction/context/response triples. We convert
+    to one-turn conversations with random system prompts.
+
+    Returns: list of (system_prompt, [(user_msg, assistant_msg)])
+    """
+    print("Loading databricks/databricks-dolly-15k ...")
+    ds = load_dataset("databricks/databricks-dolly-15k", split="train")
+
+    conversations = []
+    rng = random.Random(SEED)
+
+    for row in ds:
+        instruction = row["instruction"].strip()
+        context = row.get("context", "").strip()
+        response = row["response"].strip()
+
+        # Skip empties defensively
+        if not instruction or not response:
+            continue
+
+        # Combine instruction and context into the user message
+        if context:
+            user_msg = f"{instruction}\n\n{context}"
+        else:
+            user_msg = instruction
+
+        system_prompt = rng.choice(SYSTEM_PROMPTS)
+        turns = [(user_msg, response)]
+        conversations.append((system_prompt, turns))
+
+    print(f"  Loaded {len(conversations):,} Dolly examples")
+    return conversations
+
+
+# Room for future loaders:
+# def load_ultrachat(): ...
+# def load_oasst1(): ...
+
+
+def load_synthetic_chat(path="alter_ego_synthetic_clean.jsonl", repeat=8):
+    """
+    Load Gemini-generated synthetic conversation pairs.
+
+    These pairs (greetings, identity, capabilities, etc.) are rare in
+    UltraChat/OASST so we repeat them multiple times to give them weight
+    in the training mix.
+
+    Args:
+        path: JSONL file with rows {category, user, assistant}
+        repeat: how many copies of each pair to add to the training set.
+                Higher = stronger learning of these patterns. Default 5
+                means each pair is seen 5 times during 1 epoch.
+
+    Returns: list of [(user_msg, assistant_msg)] — single-turn conversations
+    """
+    if not os.path.isfile(path):
+        print(f"  Synthetic data file not found at {path} — skipping.")
+        return []
+
+    print(f"Loading synthetic chat data from {path} ...")
+    rows = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            row = json.loads(line)
+            rows.append(row)
+
+    print(f"  Loaded {len(rows)} unique synthetic pairs")
+    print(f"  Replicating {repeat}x for training weight = {len(rows) * repeat:,} examples")
+
+    # Convert to (system_prompt, turns) format. We use a NEUTRAL system prompt
+    # for these so the model learns the persona is intrinsic, not prompt-dependent.
+    # We don't randomize over SYSTEM_PROMPTS here — these pairs deserve to be
+    # paired with simple, consistent framing so the model learns "this is just
+    # how Alter Ego talks."
+    conversations = []
+    rng = random.Random(SEED + 100)
+    for row in rows:
+        # Mix of system prompts: half generic ("You are Alter Ego."), half
+        # from our normal pool. This lets the model generalize across system
+        # prompt variations while strongly anchoring the basic identity.
+        for _ in range(repeat):
+            if rng.random() < 0.5:
+                sys_prompt = "You are Alter Ego."
+            else:
+                sys_prompt = rng.choice(SYSTEM_PROMPTS)
+            turns = [(row["user"], row["assistant"])]
+            conversations.append((sys_prompt, turns))
+
+    return conversations
+
+
+def _rotate_system_prompts(conversations, seed=SEED):
+    """Assign a random system prompt to each conversation."""
+    rng = random.Random(seed)
+    return [(rng.choice(SYSTEM_PROMPTS), turns) for turns in conversations]
+
+
+def _truncate_turns_to_fit(enc, system_prompt, turns, max_tokens=SEQ_LEN):
+    """
+    Drop trailing turns until the rendered conversation fits in max_tokens.
+
+    Returns trimmed turns, or None if even the first turn alone won't fit.
+    """
+    for n in range(len(turns), 0, -1):
+        trial_turns = turns[:n]
+        rendered, _ = render_conversation(enc, system_prompt, trial_turns)
+        if len(rendered) <= max_tokens:
+            return trial_turns
+    return None
+
+
+def load_ultrachat(split="train_sft", max_conversations=None):
+    """
+    Load UltraChat 200K and extract multi-turn conversations.
+
+    Args:
+        split: 'train_sft' (207K convs) or 'test_sft' (23K, used for val)
+        max_conversations: take at most this many (after loading all). None = all.
+
+    Returns: list of list[(user_msg, assistant_msg)] — raw, no system prompt yet
+    """
+    print(f"Loading HuggingFaceH4/ultrachat_200k split={split} ...")
+    ds = load_dataset("HuggingFaceH4/ultrachat_200k", split=split)
+
+    conversations = []
+    for row in ds:
+        messages = row["messages"]
+        # Walk messages pairwise: [user, assistant, user, assistant, ...]
+        turns = []
+        i = 0
+        while i + 1 < len(messages):
+            u = messages[i]
+            a = messages[i + 1]
+            if u["role"] == "user" and a["role"] == "assistant":
+                turns.append((u["content"].strip(), a["content"].strip()))
+                i += 2
+            else:
+                # Malformed — skip this conversation entirely
+                turns = []
+                break
+        if turns:
+            conversations.append(turns)
+
+    print(f"  Parsed {len(conversations):,} UltraChat conversations from {split}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 3]  # oversample, filters will cut
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+def load_oasst1(max_conversations=None, require_all_rank_zero=True):
+    """
+    Load OASST1 and linearize conversation trees.
+
+    Strategy: for each conversation tree, walk from root to the best leaf.
+    'Best' = the leaf whose path has the lowest sum of ranks (rank 0 = best).
+
+    Args:
+        max_conversations: subsample target
+        require_all_rank_zero: only keep paths where every assistant rank is 0
+
+    Returns: list of list[(user_msg, assistant_msg)]
+    """
+    print("Loading OpenAssistant/oasst1 ...")
+    ds = load_dataset("OpenAssistant/oasst1", split="train")
+
+    # Build message lookup and tree structure
+    print("  Building tree structure ...")
+    messages = {}  # message_id -> row
+    children = {}  # parent_id -> [message_id]
+    roots = []
+
+    for row in ds:
+        mid = row["message_id"]
+        pid = row.get("parent_id")
+        messages[mid] = row
+        if pid is None:
+            roots.append(mid)
+        else:
+            children.setdefault(pid, []).append(mid)
+
+    print(f"  Found {len(messages):,} messages, {len(roots):,} trees")
+
+    # Walk each tree to find best linear path
+    conversations = []
+    drop_reasons = {"non_english_lang": 0, "bad_structure": 0,
+                    "rank_filter": 0, "ok": 0}
+
+    for root_id in roots:
+        root = messages[root_id]
+        # Root must be a prompter message in English
+        if root["role"] != "prompter":
+            drop_reasons["bad_structure"] += 1
+            continue
+        if root.get("lang") != "en":
+            drop_reasons["non_english_lang"] += 1
+            continue
+
+        # Walk greedy best path: at each branch, pick the child with the lowest rank
+        path = [root_id]
+        current = root_id
+        rank_sum = 0
+        bad_rank = False
+
+        while True:
+            kids = children.get(current, [])
+            if not kids:
+                break
+            # For assistant responses, sort by rank ascending (0 is best)
+            # rank can be None for some messages; treat None as worst
+            kids_sorted = sorted(
+                kids,
+                key=lambda mid: (messages[mid].get("rank") if messages[mid].get("rank") is not None else 999)
+            )
+            best_kid_id = kids_sorted[0]
+            best_kid = messages[best_kid_id]
+
+            # Track rank for assistant turns
+            if best_kid["role"] == "assistant":
+                r = best_kid.get("rank")
+                if r is None or r > 0:
+                    bad_rank = True
+                if r is not None:
+                    rank_sum += r
+
+            path.append(best_kid_id)
+            current = best_kid_id
+
+        if require_all_rank_zero and bad_rank:
+            drop_reasons["rank_filter"] += 1
+            continue
+
+        # Convert path to (user, assistant) turns
+        turns = []
+        i = 0
+        structure_ok = True
+        while i + 1 < len(path):
+            u_msg = messages[path[i]]
+            a_msg = messages[path[i + 1]]
+            if u_msg["role"] != "prompter" or a_msg["role"] != "assistant":
+                structure_ok = False
+                break
+            turns.append((u_msg["text"].strip(), a_msg["text"].strip()))
+            i += 2
+
+        if not structure_ok or not turns:
+            drop_reasons["bad_structure"] += 1
+            continue
+
+        conversations.append(turns)
+        drop_reasons["ok"] += 1
+
+    print(f"\n  OASST tree walk results:")
+    for reason, count in drop_reasons.items():
+        print(f"    {reason:>20}: {count:>6}")
+
+    if max_conversations is not None and len(conversations) > max_conversations:
+        rng = random.Random(SEED)
+        rng.shuffle(conversations)
+        conversations = conversations[:max_conversations * 2]  # oversample for filters
+        print(f"  Oversampled to {len(conversations):,} (filters will reduce to ~{max_conversations})")
+
+    return conversations
+
+
+# ─────────────────────────────────────────────────────────────
+# Quality filtering
+# ─────────────────────────────────────────────────────────────
+
+def ascii_ratio(text):
+    """Fraction of characters that are ASCII. Crude English detector."""
+    if not text:
+        return 1.0
+    ascii_count = sum(1 for c in text if ord(c) < 128)
+    return ascii_count / len(text)
+
+
+def is_likely_english(text):
+    """True if a message is probably English (ASCII-ratio based)."""
+    return ascii_ratio(text) >= MIN_ASCII_RATIO
+
+
+def starts_with_refusal_pattern(text):
+    """
+    Check if an assistant response starts with a known AI-disclaimer pattern.
+    Match is on the first ~80 characters, case-insensitive, stripped of leading whitespace.
+    """
+    if not text:
+        return False
+    head = text.strip().lower()[:80]
+    for pat in REFUSAL_PATTERNS:
+        if head.startswith(pat):
+            return True
+    return False
+
+
+def passes_quality_filter(enc, system_prompt, turns, strict=False, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Returns (ok, reason).
+
+    Args:
+        strict: if True, apply production filters (language, absolute assistant
+                minimum length). If False, only apply basic filters (Dolly default).
+        filter_refusals: if True, drop conversations whose assistant response
+                starts with a known "as an AI" / "I am not capable" pattern.
+        max_assistant_tokens: if set, drop conversations whose assistant response
+                exceeds this many tokens in ANY turn. None = no cap.
+    """
+    if len(turns) < 1:
+        return False, "too_few_turns"
+
+    for user_msg, assistant_msg in turns:
+        if not user_msg.strip() or not assistant_msg.strip():
+            return False, "empty_msg"
+
+        # Length check in tokens
+        u_toks = len(encode_plain(enc, user_msg))
+        a_toks = len(encode_plain(enc, assistant_msg))
+        if u_toks < MIN_MSG_TOKENS or a_toks < MIN_MSG_TOKENS:
+            return False, "msg_too_short"
+
+        if strict:
+            # Language filter
+            if not is_likely_english(user_msg) or not is_likely_english(assistant_msg):
+                return False, "non_english"
+
+            # Absolute minimum length on assistant response.
+            # We use an absolute floor rather than a ratio because tying output
+            # length to input length would force the model to pad short
+            # answers to long (e.g. RAG) prompts — teaching it to yap.
+            if a_toks < MIN_ASSISTANT_TOKENS:
+                return False, "assistant_too_short"
+
+        # Optional upper bound (CLI flag). Applies regardless of strict mode
+        # so Dolly could also use it if requested.
+        if max_assistant_tokens is not None and a_toks > max_assistant_tokens:
+            return False, "assistant_too_long"
+
+        if filter_refusals and starts_with_refusal_pattern(assistant_msg):
+            return False, "refusal_pattern"
+
+    return True, "ok"
+
+
+def fits_in_sequence(tokens):
+    """Check that the rendered conversation fits in our seq length."""
+    return len(tokens) <= SEQ_LEN
+
+
+# ─────────────────────────────────────────────────────────────
+# Main pipeline
+# ─────────────────────────────────────────────────────────────
+
+def _filter_and_render(enc, conversations_with_system, strict, cap=None, truncate_to_fit=False, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Apply quality filters and render to (tokens, mask).
+
+    Args:
+        conversations_with_system: list of (system_prompt, turns)
+        strict: use production filters (language, length ratio)
+        cap: stop once we have this many rendered examples (None = no cap)
+        truncate_to_fit: if True, drop trailing turns to fit in SEQ_LEN
+                         instead of dropping the whole conversation
+        filter_refusals: if True, drop conversations with AI-disclaimer openers
+        max_assistant_tokens: if set, drop convs with any assistant msg > this
+
+    Returns: (rendered_list, drop_reasons_dict)
+    """
+    rendered = []
+    drop_reasons = {
+        "too_few_turns": 0, "empty_msg": 0, "msg_too_short": 0,
+        "non_english": 0, "assistant_too_short": 0,
+        "assistant_too_long": 0, "refusal_pattern": 0,
+        "too_long": 0, "ok": 0,
+    }
+
+    for system_prompt, turns in tqdm(conversations_with_system, desc="Rendering"):
+        if cap is not None and len(rendered) >= cap:
+            break
+
+        ok, reason = passes_quality_filter(
+            enc, system_prompt, turns,
+            strict=strict, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+        if not ok:
+            drop_reasons[reason] = drop_reasons.get(reason, 0) + 1
+            continue
+
+        # Try to fit in sequence length
+        if truncate_to_fit:
+            fit_turns = _truncate_turns_to_fit(enc, system_prompt, turns)
+            if fit_turns is None:
+                drop_reasons["too_long"] += 1
+                continue
+            turns_to_render = fit_turns
+        else:
+            turns_to_render = turns
+
+        tokens, mask = render_conversation(enc, system_prompt, turns_to_render)
+        if len(tokens) > SEQ_LEN:
+            drop_reasons["too_long"] += 1
+            continue
+
+        # Must have at least some loss
+        if sum(mask) == 0:
+            drop_reasons["msg_too_short"] += 1
+            continue
+
+        rendered.append((tokens, mask))
+        drop_reasons["ok"] += 1
+
+    return rendered, drop_reasons
+
+
+def _build_train_val_for_dolly(enc, max_assistant_tokens=None):
+    """Dolly: load all, apply basic filters, 95/5 random split."""
+    raw = load_dolly()  # already has system prompts assigned
+    rendered, drops = _filter_and_render(
+        enc, raw, strict=False,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_ultrachat(enc, filter_refusals=False, max_assistant_tokens=None):
+    """UltraChat alone: train_sft for training, test_sft for validation."""
+    train_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+
+    train_convs = _rotate_system_prompts(train_raw, seed=SEED)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+
+    train_rendered, train_drops = _filter_and_render(
+        enc, train_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    return train_rendered, val_rendered, {"train": train_drops, "val": val_drops}
+
+
+def _build_train_val_for_oasst(enc, filter_refusals=False, max_assistant_tokens=None):
+    """OASST alone: load tree, linearize, 95/5 random split."""
+    raw = load_oasst1(max_conversations=OASST_TARGET)
+    convs = _rotate_system_prompts(raw, seed=SEED)
+    rendered, drops = _filter_and_render(
+        enc, convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    rng = random.Random(SEED)
+    rng.shuffle(rendered)
+    n_val = max(1, int(len(rendered) * VAL_FRACTION))
+    return rendered[n_val:], rendered[:n_val], {"total": drops}
+
+
+def _build_train_val_for_prod(enc, filter_refusals=False, max_assistant_tokens=None):
+    """
+    Production: UltraChat train_sft + OASST for training,
+                UltraChat test_sft for validation.
+    """
+    # Training data from both sources
+    print("\n[1/3] Loading UltraChat for training ...")
+    uc_raw = load_ultrachat(split="train_sft", max_conversations=ULTRACHAT_TARGET)
+    uc_convs = _rotate_system_prompts(uc_raw, seed=SEED)
+    uc_rendered, uc_drops = _filter_and_render(
+        enc, uc_convs, strict=True, cap=ULTRACHAT_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  UltraChat accepted: {len(uc_rendered):,}")
+
+    print("\n[2/3] Loading OASST for training ...")
+    oasst_raw = load_oasst1(max_conversations=OASST_TARGET)
+    oasst_convs = _rotate_system_prompts(oasst_raw, seed=SEED + 2)
+    oasst_rendered, oasst_drops = _filter_and_render(
+        enc, oasst_convs, strict=True, cap=OASST_TARGET, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  OASST accepted: {len(oasst_rendered):,}")
+
+    # ── Synthetic conversational pairs (greetings, identity, capabilities) ──
+    # These are ABSENT from UltraChat/OASST so we add them explicitly. They
+    # train the model to handle short casual inputs and to know its identity.
+    # Loaded only if alter_ego_synthetic_clean.jsonl exists.
+    print("\n[2.5/3] Loading synthetic conversational pairs ...")
+    synth_raw = load_synthetic_chat(repeat=8)  # each pair sees model 8 times
+    synth_drops = {"ok": 0}
+    synth_rendered = []
+    if synth_raw:
+        # Synthetic pairs use 'strict=False' since they're already curated by
+        # us and don't need the language/refusal/length-floor checks.
+        synth_rendered, synth_drops = _filter_and_render(
+            enc, synth_raw, strict=False, truncate_to_fit=True,
+            filter_refusals=False,            # we wrote these, no refusals
+            max_assistant_tokens=None,        # already short by design
+        )
+        print(f"  Synthetic accepted: {len(synth_rendered):,}")
+
+    # Combine and shuffle
+    train_rendered = uc_rendered + oasst_rendered + synth_rendered
+    rng = random.Random(SEED)
+    rng.shuffle(train_rendered)
+    print(f"\n  Combined training set: {len(train_rendered):,} conversations")
+
+    # Validation from UltraChat test_sft only (clean, no OASST noise)
+    print("\n[3/3] Loading UltraChat test_sft for validation ...")
+    val_raw = load_ultrachat(split="test_sft", max_conversations=500)
+    val_convs = _rotate_system_prompts(val_raw, seed=SEED + 1)
+    val_rendered, val_drops = _filter_and_render(
+        enc, val_convs, strict=True, cap=500, truncate_to_fit=True,
+        filter_refusals=filter_refusals,
+        max_assistant_tokens=max_assistant_tokens,
+    )
+    print(f"  Val accepted: {len(val_rendered):,}")
+
+    return train_rendered, val_rendered, {
+        "ultrachat_train": uc_drops,
+        "oasst_train": oasst_drops,
+        "synthetic_train": synth_drops,
+        "val": val_drops,
+        "sources": {
+            "ultrachat": len(uc_rendered),
+            "oasst": len(oasst_rendered),
+            "synthetic": len(synth_rendered),
+        },
+    }
+
+
+def process_dataset(dataset_name, output_dir, filter_refusals=False, max_assistant_tokens=None):
+    enc = get_tokenizer()
+
+    # Route to the appropriate builder
+    print(f"\n{'='*70}")
+    print(f"Processing dataset: {dataset_name}")
+    print(f"  filter_refusals:        {filter_refusals}")
+    print(f"  max_assistant_tokens:   {max_assistant_tokens}")
+    print(f"{'='*70}")
+
+    if dataset_name == "dolly":
+        # Dolly: refusal filter not wired (single-turn, rare patterns there)
+        train_data, val_data, filter_info = _build_train_val_for_dolly(
+            enc, max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "ultrachat":
+        train_data, val_data, filter_info = _build_train_val_for_ultrachat(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "oasst":
+        train_data, val_data, filter_info = _build_train_val_for_oasst(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    elif dataset_name == "prod":
+        train_data, val_data, filter_info = _build_train_val_for_prod(
+            enc, filter_refusals=filter_refusals,
+            max_assistant_tokens=max_assistant_tokens,
+        )
+    else:
+        raise ValueError(f"Unknown dataset: {dataset_name}")
+
+    if len(train_data) == 0:
+        raise RuntimeError("No training examples survived filtering. Check your data.")
+    if len(val_data) == 0:
+        raise RuntimeError("No validation examples survived filtering.")
+
+    print(f"\n  Train: {len(train_data):,} examples")
+    print(f"  Val:   {len(val_data):,} examples")
+
+    # Pad and convert to arrays
+    print("\nPadding and converting to arrays ...")
+
+    def to_arrays(data):
+        n = len(data)
+        tokens_arr = np.full((n, SEQ_LEN), PAD_TOKEN_ID, dtype=np.uint32)
+        mask_arr = np.zeros((n, SEQ_LEN), dtype=np.uint8)
+        for i, (toks, msk) in enumerate(data):
+            toks_padded, msk_padded = pad_to_seq_len(toks, msk)
+            tokens_arr[i] = toks_padded
+            mask_arr[i] = msk_padded
+        return tokens_arr, mask_arr
+
+    train_tokens, train_mask = to_arrays(train_data)
+    val_tokens, val_mask = to_arrays(val_data)
+
+    # Save
+    os.makedirs(output_dir, exist_ok=True)
+    np.save(os.path.join(output_dir, "sft_train.npy"), train_tokens)
+    np.save(os.path.join(output_dir, "sft_train_mask.npy"), train_mask)
+    np.save(os.path.join(output_dir, "sft_val.npy"), val_tokens)
+    np.save(os.path.join(output_dir, "sft_val_mask.npy"), val_mask)
+
+    # Metadata
+    def stats(tokens_arr, mask_arr):
+        real_lens = (tokens_arr != PAD_TOKEN_ID).sum(axis=1)
+        loss_fractions = mask_arr.sum(axis=1) / real_lens.clip(min=1)
+        return {
+            "num_examples": int(tokens_arr.shape[0]),
+            "total_tokens": int(real_lens.sum()),
+            "total_loss_tokens": int(mask_arr.sum()),
+            "avg_length": float(real_lens.mean()),
+            "median_length": float(np.median(real_lens)),
+            "min_length": int(real_lens.min()),
+            "max_length": int(real_lens.max()),
+            "avg_loss_fraction": float(loss_fractions.mean()),
+        }
+
+    metadata = {
+        "dataset": dataset_name,
+        "seq_len": SEQ_LEN,
+        "pad_token_id": PAD_TOKEN_ID,
+        "im_start_id": IM_START_ID,
+        "im_end_id": IM_END_ID,
+        "eot_id": EOT_ID,
+        "system_prompts": SYSTEM_PROMPTS,
+        "filter_refusals": filter_refusals,
+        "refusal_patterns": REFUSAL_PATTERNS if filter_refusals else None,
+        "max_assistant_tokens": max_assistant_tokens,
+        "filter_results": filter_info,
+        "train": stats(train_tokens, train_mask),
+        "val": stats(val_tokens, val_mask),
+    }
+
+    with open(os.path.join(output_dir, "sft_metadata.json"), "w") as f:
+        json.dump(metadata, f, indent=2)
+
+    print(f"\n  Wrote arrays and metadata to {output_dir}/")
+
+    # Preview
+    print("\n" + "=" * 70)
+    print("PREVIEW: 3 random examples (abbreviated)")
+    print("=" * 70)
+    rng = random.Random(SEED)
+    preview_indices = rng.sample(range(len(train_data)), k=min(3, len(train_data)))
+    for idx in preview_indices:
+        tokens, mask = train_data[idx]
+        preview_example(enc, tokens, mask, max_tokens=80)
+
+    print("\nDone. Next: run test_preprocessing.py to verify correctness.")
+    return metadata
+
+
+def preview_example(enc, tokens, mask, max_tokens=80):
+    """Print a human-readable preview showing tokens with their mask values."""
+    print("\n" + "-" * 70)
+    print(f"Total tokens: {len(tokens)}, loss tokens: {sum(mask)}")
+    print(f"First {min(max_tokens, len(tokens))} tokens:")
+    print(f"{'IDX':>4} {'TOK_ID':>7} {'MASK':>4}  TEXT")
+    for i, (tok, m) in enumerate(zip(tokens[:max_tokens], mask[:max_tokens])):
+        try:
+            text = enc.decode([tok])
+        except Exception:
+            text = "<decode-error>"
+        marker = "◀LOSS" if m else ""
+        # Escape newlines for readability
+        text_display = repr(text)[1:-1][:40]
+        print(f"{i:>4} {tok:>7} {m:>4}  {text_display} {marker}")
+    if len(tokens) > max_tokens:
+        print(f"  ... ({len(tokens) - max_tokens} more tokens)")
+
+
+# ─────────────────────────────────────────────────────────────
+# CLI entry point
+# ─────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--dataset",
+        choices=["dolly", "ultrachat", "oasst", "prod"],
+        default="dolly",
+        help="Which dataset to process. 'prod' = UltraChat + OASST mix (recommended for real run)."
+    )
+    parser.add_argument(
+        "--output", type=str, default=None,
+        help="Output directory (default: sft_data_<dataset>[_clean][_maxN])"
+    )
+    parser.add_argument(
+        "--filter-refusals", action="store_true",
+        help="Drop conversations whose assistant response starts with 'as an AI', "
+             "'I am not capable', etc. Does not apply to Dolly."
+    )
+    parser.add_argument(
+        "--max-assistant-tokens", type=int, default=None,
+        help="If set, drop conversations where any assistant message exceeds "
+             "this many tokens. Biases training toward shorter responses. "
+             "Try 700 (drops ~50%% UltraChat, teaches concise style) or "
+             "1000 (drops ~30%%, mild outlier filter)."
+    )
+    args = parser.parse_args()
+
+    # Default output dir reflects what filters were applied so runs don't collide.
+    if args.output:
+        output_dir = args.output
+    else:
+        suffix = ""
+        if args.filter_refusals:
+            suffix += "_clean"
+        if args.max_assistant_tokens is not None:
+            suffix += f"_max{args.max_assistant_tokens}"
+        output_dir = f"sft_data_{args.dataset}{suffix}"
+
+    process_dataset(
+        args.dataset, output_dir,
+        filter_refusals=args.filter_refusals,
+        max_assistant_tokens=args.max_assistant_tokens,
+    )

workspace/Alter_Ego/run_sft_training.sh

@@ -0,0 +1,84 @@
+#!/bin/bash
+# Launch SFT training with stdout captured to a log file.
+#
+# Usage:
+#   ./run_sft.sh                          # run with defaults from trainsft.py
+#   DATA_DIR=sft_data_prod ./run_sft.sh   # override SFT_DATA_DIR via env var
+#   CKPT=llme_model_step_19072.pt ./run_sft.sh   # override pretraining checkpoint
+#
+# Notes:
+# - Unlike pretraining, SFT does not currently support resume from a partial
+#   SFT checkpoint. If a run dies, restart from the original pretraining
+#   checkpoint. The whole SFT run is ~1 hour so this is not painful.
+# - All optional env vars below override the constants set inside trainsft.py.
+
+cd /workspace/Alter_Ego
+
+# Optional overrides via env vars (each one falls back to the trainsft.py default)
+DATA_DIR=${DATA_DIR:-}
+CKPT=${CKPT:-}
+
+TIMESTAMP=$(date +%Y-%m-%d_%H-%M-%S)
+STDOUT_LOG="sft_stdout_${TIMESTAMP}.log"
+
+echo "============================================================"
+echo "Launching SFT training"
+echo "  timestamp:  $TIMESTAMP"
+echo "  stdout ->   $STDOUT_LOG"
+if [ -n "$DATA_DIR" ]; then
+    echo "  DATA_DIR =  $DATA_DIR  (overriding trainsft.py default)"
+fi
+if [ -n "$CKPT" ]; then
+    echo "  CKPT =      $CKPT  (overriding trainsft.py default)"
+fi
+echo "============================================================"
+
+# Pre-flight: confirm key files exist on the pod
+MISSING=""
+[ -f trainsft.py ] || MISSING="$MISSING trainsft.py"
+[ -f prep_sft_data.py ] || MISSING="$MISSING prep_sft_data.py"
+
+# Determine which checkpoint and data dir to validate
+EFFECTIVE_CKPT=${CKPT:-$(grep "^PRETRAIN_CHECKPOINT" trainsft.py | head -1 | cut -d"'" -f2)}
+EFFECTIVE_DATA=${DATA_DIR:-$(grep "^SFT_DATA_DIR" trainsft.py | head -1 | cut -d"'" -f2)}
+
+[ -f "$EFFECTIVE_CKPT" ] || MISSING="$MISSING $EFFECTIVE_CKPT"
+[ -d "$EFFECTIVE_DATA" ] || MISSING="$MISSING $EFFECTIVE_DATA/"
+
+if [ -n "$MISSING" ]; then
+    echo "ERROR: missing required files:$MISSING"
+    echo "Aborting — fix the above and re-run."
+    exit 1
+fi
+
+echo "Pre-flight OK:"
+echo "  pretraining checkpoint: $EFFECTIVE_CKPT"
+echo "  SFT data directory:     $EFFECTIVE_DATA"
+echo "---"
+
+# Export env vars so trainsft.py sees them (only used if you've added env-var
+# support inside trainsft.py; harmless otherwise)
+export SFT_DATA_DIR="$DATA_DIR"
+export PRETRAIN_CHECKPOINT="$CKPT"
+
+# Launch
+python -u trainsft.py 2>&1 | tee "$STDOUT_LOG"
+
+EXIT_CODE=${PIPESTATUS[0]}
+echo "---"
+echo "trainsft.py exit code: $EXIT_CODE"
+
+# Sync logs and checkpoints to HF on success
+if [ "$EXIT_CODE" = "0" ]; then
+    echo "Run completed. Syncing artifacts to HuggingFace ..."
+    if command -v hf &>/dev/null; then
+        hf upload jbomdev/Alter-Ego-350m sft_logs/ sft_logs/ 2>&1 || \
+            echo "  (sft_logs/ upload failed, retry manually)"
+        hf upload jbomdev/Alter-Ego-350m sft_checkpoints/ sft_checkpoints/ 2>&1 || \
+            echo "  (sft_checkpoints/ upload failed, retry manually)"
+    else
+        echo "  huggingface-cli not found; skipping auto-upload."
+    fi
+fi
+
+exit $EXIT_CODE

workspace/Alter_Ego/run_training.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+# Launch training with stdout captured to a log file
+
+cd /workspace/Alter_Ego
+
+# If RESUME=1, will resume from latest checkpoint
+RESUME=${RESUME:-0}
+
+TIMESTAMP=$(date +%Y-%m-%d_%H-%M-%S)
+STDOUT_LOG="training_stdout_${TIMESTAMP}.log"
+
+echo "Starting training (RESUME=$RESUME)"
+echo "stdout -> $STDOUT_LOG"
+echo "---"
+
+RESUME=$RESUME python -u trainv2.py 2>&1 | tee "$STDOUT_LOG"

workspace/Alter_Ego/sft_checkpoints/alter.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9af281ab89a3be0df12f755ea72402d7b3c08371d0d2d5be4ac9afacecb3892c
+size 4480339414

workspace/Alter_Ego/sft_checkpoints/llme_sft_step_200.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:221e3655ec76125993f0a61b92a36adc8a975b098cc1182145a924a5e9c19b09
+size 4480339414

workspace/Alter_Ego/sft_checkpoints/llme_sft_step_222.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27eccaf6a54e9c32e9239bcf713267654cd0b7aac42606dccd536417fbe85f06
+size 4480339414

workspace/Alter_Ego/sft_checkpoints/llme_sft_step_249.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24a3db2d79f52df6b87d7de5378708bff9502d7384a5e00a8010d5a51d0b315e
+size 4480339414

workspace/Alter_Ego/sft_checkpoints_3ep_failed/llme_sft_step_311.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac6085883b618ab85fe750be7f52d66383bd48e893de9e421dd0678b08e99f81
+size 4480339414

workspace/Alter_Ego/sft_checkpoints_3ep_failed/llme_sft_step_400.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebed8c96af8ffca9f89e224734bae9f33d5c233ac76cf1203ffea02a09bcd210
+size 4480339414

workspace/Alter_Ego/sft_checkpoints_3ep_failed/llme_sft_step_600.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:114ffda81b11f709eb0c980518db3677e50067430b16dbfcec35a1d24d438793
+size 4480339414

workspace/Alter_Ego/sft_checkpoints_3ep_failed/llme_sft_step_668.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2be11af8d48726979b6dccca3a92cbd3b89a341c4eb04cad52e03150dce7de
+size 4480339414

workspace/Alter_Ego/sft_data_dolly/sft_metadata.json

@@ -0,0 +1,56 @@
+{
+  "dataset": "dolly",
+  "seq_len": 2048,
+  "pad_token_id": 100257,
+  "im_start_id": 100277,
+  "im_end_id": 100278,
+  "eot_id": 100257,
+  "system_prompts": [
+    "You are Alter Ego. You enjoy explaining things clearly. Speak casually and use contractions.",
+    "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out.",
+    "You are Alter Ego. Be friendly and get to the point. Keep it natural.",
+    "You are Alter Ego. Explain things simply and conversationally.",
+    "You are Alter Ego. Be warm, relaxed, and conversational.",
+    "You are Alter Ego. Share what you know in a friendly, easy-to-understand way.",
+    "You are Alter Ego. You're a clever and approachable assistant. Keep it casual.",
+    "You are Alter Ego, a helpful and friendly AI.",
+    "You are Alter Ego. Provide clear, accurate answers.",
+    "You are Alter Ego. You find learning interesting and enjoy breaking down complex topics into simple terms."
+  ],
+  "filter_refusals": false,
+  "refusal_patterns": null,
+  "max_assistant_tokens": null,
+  "filter_results": {
+    "total": {
+      "too_few_turns": 0,
+      "empty_msg": 0,
+      "msg_too_short": 988,
+      "non_english": 0,
+      "assistant_too_short": 0,
+      "assistant_too_long": 0,
+      "refusal_pattern": 0,
+      "too_long": 0,
+      "ok": 14023
+    }
+  },
+  "train": {
+    "num_examples": 13322,
+    "total_tokens": 2739942,
+    "total_loss_tokens": 1108731,
+    "avg_length": 205.6704698994145,
+    "median_length": 138.0,
+    "min_length": 37,
+    "max_length": 2048,
+    "avg_loss_fraction": 0.4247286763562959
+  },
+  "val": {
+    "num_examples": 701,
+    "total_tokens": 141500,
+    "total_loss_tokens": 55772,
+    "avg_length": 201.85449358059915,
+    "median_length": 133.0,
+    "min_length": 42,
+    "max_length": 1969,
+    "avg_loss_fraction": 0.41722495059473474
+  }
+}

workspace/Alter_Ego/sft_data_dolly/sft_train.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:712797aca4afedea95ac3c102ddee9695fe4b7c5f399b0c389885a5195c72da6
+size 109133952

workspace/Alter_Ego/sft_data_dolly/sft_train_mask.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fc18c0475730ff2c40ab225590b51f6a54c597fdb4ba890a733456d4e40b71f
+size 27283584

workspace/Alter_Ego/sft_data_dolly/sft_val.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7a15991d1762349803bed72aaa8b531474a37f3c5fd47a042e9005bdf558c74
+size 5742720

workspace/Alter_Ego/sft_data_dolly/sft_val_mask.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff90fdd8aa40be38cc6c8e3794cded2d8888cdfb22e581208d727221785040fb
+size 1435776

workspace/Alter_Ego/sft_data_prod_clean/sft_metadata.json

@@ -0,0 +1,127 @@
+{
+  "dataset": "prod",
+  "seq_len": 2048,
+  "pad_token_id": 100257,
+  "im_start_id": 100277,
+  "im_end_id": 100278,
+  "eot_id": 100257,
+  "system_prompts": [
+    "You are Alter Ego. You enjoy explaining things clearly. Speak casually and use contractions.",
+    "You are Alter Ego. Answer directly and use everyday language, like a smart friend helping out.",
+    "You are Alter Ego. Be friendly and get to the point. Keep it natural.",
+    "You are Alter Ego. Explain things simply and conversationally.",
+    "You are Alter Ego. Be warm, relaxed, and conversational.",
+    "You are Alter Ego. Share what you know in a friendly, easy-to-understand way.",
+    "You are Alter Ego. You're a clever and approachable assistant. Keep it casual.",
+    "You are Alter Ego, a helpful and friendly AI.",
+    "You are Alter Ego. Provide clear, accurate answers.",
+    "You are Alter Ego. You find learning interesting and enjoy breaking down complex topics into simple terms."
+  ],
+  "filter_refusals": true,
+  "refusal_patterns": [
+    "as an ai",
+    "as a language model",
+    "as an ai language model",
+    "as an ai assistant",
+    "i am an ai",
+    "i'm an ai",
+    "i am just an ai",
+    "i'm just an ai",
+    "i am not capable",
+    "i'm not capable",
+    "i am not able",
+    "i'm not able",
+    "i don't have the ability",
+    "i do not have the ability",
+    "i don't have access",
+    "i do not have access",
+    "i am unable to",
+    "i'm unable to",
+    "i cannot browse",
+    "i can't browse",
+    "i cannot provide",
+    "i can't provide personal",
+    "i don't have personal",
+    "i do not have personal",
+    "i don't have feelings",
+    "i do not have feelings",
+    "i don't have emotions",
+    "i do not have emotions",
+    "i don't have opinions",
+    "i do not have opinions",
+    "as a responsible ai",
+    "as an artificial intelligence"
+  ],
+  "max_assistant_tokens": 700,
+  "filter_results": {
+    "ultrachat_train": {
+      "too_few_turns": 0,
+      "empty_msg": 2,
+      "msg_too_short": 64,
+      "non_english": 10,
+      "assistant_too_short": 78,
+      "assistant_too_long": 3566,
+      "refusal_pattern": 1800,
+      "too_long": 0,
+      "ok": 25714
+    },
+    "oasst_train": {
+      "too_few_turns": 0,
+      "empty_msg": 0,
+      "msg_too_short": 144,
+      "non_english": 7,
+      "assistant_too_short": 29,
+      "assistant_too_long": 58,
+      "refusal_pattern": 70,
+      "too_long": 0,
+      "ok": 2863
+    },
+    "synthetic_train": {
+      "too_few_turns": 0,
+      "empty_msg": 0,
+      "msg_too_short": 1250,
+      "non_english": 0,
+      "assistant_too_short": 0,
+      "assistant_too_long": 0,
+      "refusal_pattern": 0,
+      "too_long": 0,
+      "ok": 2770
+    },
+    "val": {
+      "too_few_turns": 0,
+      "empty_msg": 0,
+      "msg_too_short": 0,
+      "non_english": 0,
+      "assistant_too_short": 2,
+      "assistant_too_long": 66,
+      "refusal_pattern": 39,
+      "too_long": 0,
+      "ok": 500
+    },
+    "sources": {
+      "ultrachat": 25714,
+      "oasst": 2863,
+      "synthetic": 2770
+    }
+  },
+  "train": {
+    "num_examples": 31347,
+    "total_tokens": 29372553,
+    "total_loss_tokens": 21490235,
+    "avg_length": 937.0132070054551,
+    "median_length": 920.0,
+    "min_length": 37,
+    "max_length": 2048,
+    "avg_loss_fraction": 0.6862104882242435
+  },
+  "val": {
+    "num_examples": 500,
+    "total_tokens": 542705,
+    "total_loss_tokens": 385913,
+    "avg_length": 1085.41,
+    "median_length": 1038.0,
+    "min_length": 198,
+    "max_length": 2048,
+    "avg_loss_fraction": 0.6943630272027509
+  }
+}

workspace/Alter_Ego/sft_data_prod_clean/sft_train.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6af5ad20ba08ba0cba681f2a67fb6d2270024f4a1f83a19e687c218f07ed0a2
+size 256794752

workspace/Alter_Ego/sft_data_prod_clean/sft_train_mask.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c9a7c467cc511b1113875466a51d26a90c713937fcc522ed628e54924229efb
+size 64198784

workspace/Alter_Ego/sft_data_prod_clean/sft_val.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f0c06c2e8312bf23a0314506bf4d60b75851461832195f2a2d3406a16cd7a2a
+size 4096128

workspace/Alter_Ego/sft_data_prod_clean/sft_val_mask.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da4f1fefb4858f1419400bda5e1b4fdece4661bb799af10b0bebbf8d0d2ed9b4
+size 1024128

workspace/Alter_Ego/sft_data_stage2_persona/sft_metadata.json

@@ -0,0 +1,13 @@
+{
+  "dataset": "stage2_persona_only",
+  "train_shape": [
+    3819,
+    2048
+  ],
+  "val_shape": [
+    201,
+    2048
+  ],
+  "seq_len": 2048,
+  "pad_token_id": 100257
+}

workspace/Alter_Ego/sft_data_stage2_persona/sft_train.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7266b547ece0246438342ca5ba3dfa0762e3aa8c50261364fada2f80c729497
+size 31285376

workspace/Alter_Ego/sft_data_stage2_persona/sft_train_mask.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eb2e332b178edb03c40a358b0d7866ff0eb7234441b57bb77ae6209a7a83232
+size 7821440

workspace/Alter_Ego/sft_data_stage2_persona/sft_val.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98ba3812609ed27cbf74c84ceeb6dcbc7ca0d3500d90a80b8aa9f9b4b631b422
+size 1646720

workspace/Alter_Ego/sft_data_stage2_persona/sft_val_mask.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0387c84fb790937bc43e74d5847a431aa73922a3d8a2a6cf3789a97aa915b214
+size 411776

workspace/Alter_Ego/sft_logs/SAMPLES_2026-04-25_10-22-39.txt

@@ -0,0 +1,17 @@
+
+======================================================================
+Step 0  (lr=4.00e-07)
+Prompt: What is the capital of France?
+Reply:  It is the year 2033. The National Geographic Society in the U.S. has launched a new website to raise awareness about the importance of the environment and to inspire the next generation of global citizens. The online resource is called
+“The Ecologist,” a new interactive website that will provide students and the general public with the opportunity to learn about how humans have affected the planet.
+The website will provide students with the opportunity to study a wide variety of information related to the environment and the effects of
+
+======================================================================
+Step 100  (lr=1.65e-05)
+Prompt: What is the capital of France?
+Reply:  The capital of France is Paris. This is the largest city in the country and is located in the northwestern part of the country. It is home to over 10 million people and is home to many cultural centers and museums. The city has a significant population of people with a diverse population of people from different regions and backgrounds. The capital city is also home to many cultural centers and museums, including the Louvre (which is a museum with exhibits of the world's greatest works of art), the
+
+======================================================================
+Step 200  (lr=2.77e-06)
+Prompt: What is the capital of France?
+Reply:  The capital of France is Paris.

workspace/Alter_Ego/sft_logs/SAMPLES_2026-04-25_11-54-32.txt

@@ -0,0 +1,18 @@
+
+======================================================================
+Step 0  (lr=4.00e-07)
+Prompt: What is the capital of France?
+Reply:  It is the year 2033. The National Geographic Society in the U.S. has launched a new website to raise awareness about the importance of the environment and to inspire the next generation of global citizens. The online resource is called
+“The Ecologist,” a new interactive website that will provide students and the general public with the opportunity to learn about how humans have affected the planet.
+The website will provide students with the opportunity to study a wide variety of information related to the environment and the effects of
+
+======================================================================
+Step 100  (lr=1.74e-05)
+Prompt: What is the capital of France?
+Reply:  The capital of France is Paris.
+
+======================================================================
+Step 200  (lr=4.64e-06)
+Prompt: What is the capital of France?
+Reply:  What is the capital of the country of France?
+The capital city of France is Paris, located in the north of France.

workspace/Alter_Ego/sft_logs/SAMPLES_2026-04-25_14-39-31.txt

@@ -0,0 +1,17 @@
+
+======================================================================
+Step 0  (lr=4.00e-07)
+Prompt: What is the capital of France?
+Reply:  It is the year 2033. The National Geographic Society in the U.S. has launched a new website to raise awareness about the importance of the environment and to inspire the next generation of global citizens. The online resource is called
+“The Ecologist,” a new interactive website that will provide students and the general public with the opportunity to learn about how humans have affected the planet.
+The website will provide students with the opportunity to study a wide variety of information related to the environment and the effects of
+
+======================================================================
+Step 100  (lr=1.72e-05)
+Prompt: What is the capital of France?
+Reply:  The capital of France is Paris.
+
+======================================================================
+Step 200  (lr=4.19e-06)
+Prompt: What is the capital of France?
+Reply:  What is the name of the country?

workspace/Alter_Ego/sft_logs/SAMPLES_2026-04-25_15-29-12.txt

@@ -0,0 +1,5 @@
+
+======================================================================
+Step 0  (lr=4.00e-07)
+Prompt: What is the capital of France?
+Reply:  It's the French capital, which is the capital city of France.