Model-port identity registration request: falcon7b, pythia410m, bloom560m, xverse7b

[DarthCeltic]

Requesting stage-1 identity registration for 4 model-port candidates

Per docs/MODEL_PORTS_TRACK_PLAN.md's two-stage process, this is a
registration ask only -- no credit claimed here, and I understand this
review/approval PR (if you make one) earns no credit itself.

All 4 are genuinely distinct llama.cpp-et execution families (not seed
families, not size/quant variants of anything already registered), pinned to
an exact upstream commit (not a floating main), ungated, and permissively
licensed. Each has a working recipe with local sysemu verification (full ET
offload, correct param count, clean compute graph) already on our open PRs.

Candidate model	Execution family	Upstream repo	Pinned revision	License
falcon7b	falcon	tiiuae/falcon-7b	ec89142b67d748a1865ea4451372db8313ada0d8	Apache-2.0
pythia410m	gptneox	EleutherAI/pythia-410m	9879c9b5f8bea9051dcb0e68dff21493d67e9d4f	Apache-2.0
bloom560m	bloom	bigscience/bloom-560m	ac2ae5fab2ce3f9f40dc79b5ca9f637430d24971	BigScience RAIL-1.0
xverse7b	xverse	xverse/XVERSE-7B-Chat	60acc8c453c067b54df88be98bfdf60585ab5441	Apache-2.0

Recipes with the full provenance/conversion/verification detail for each are
already on open PRs:

• falcon7b: PR feat(llama_cpp_et): add Falcon-7B and Arcee AFM-4.5B causal-LM ports #183 (ported_models/llama_cpp_et/falcon7b_recipe.md)
• pythia410m / bloom560m: PR docs(llama_cpp_et): Pythia-410m and BLOOM-560m ports (self-converted, verified, hosting pending) #189 (self-converted -- no pre-made Q8_0 GGUF
  exists for either; recipes document a real upstream convert_hf_to_gguf.py
  bug found+fixed along the way)
• xverse7b: PR feat(llama_cpp_et): add XVERSE-7B-Chat causal-LM port #188 (ported_models/llama_cpp_et/xverse7b_recipe.md)

I know stage 1 also needs a reviewed host oracle/validation contract per new
family, not just the identity entry -- happy to draft a PPL-based contract for
each (same shape as the existing llama32_1b/smolvlm_500m contracts, since
all 4 are text-only causal LMs) for your review, if that's useful, rather than
you writing it from scratch. Let me know and I'll open that as its own PR.

Two more candidates exist (mpt7b, arch mpt/ALiBi; cohere2_7b, arch
cohere2/sliding-window) but both upstream repos are gated
(mosaicml/mpt-7b 401s the public API; CohereLabs/c4ai-command-r7b-12-2024
is gated: auto and CC-BY-NC-4.0) -- flagging honestly rather than omitting,
but leading with the 4 above since they're frictionless to reproduce.

[DarthCeltic]

Update: full roster + implementation PRs now in the compliant claim shape

Following up on the 4 above -- went back through every PR we had open on this
track and brought each into the shape docs/MODEL_PORTS_TRACK_PLAN.md
actually requires: a new standalone ported_models/<model> root (just
docs/RECIPE.md, nothing else) plus ported_models/submissions/model_ports/<model>.json,
no shared-file edits inside the claim itself. Full list, all distinct
llama.cpp-et execution families (checked against data/model-port-identities.json
and against each other -- no duplicates, no seed-family reuse):

Candidate	Execution family (proposed identity_id)	Upstream repo	Pinned revision	License	PR
gpt2	gpt2	openai-community/gpt2	607a30d783dfa663caf39e06633721c8d4cfcd7e	MIT	#176
gemma2_2b	gemma2	google/gemma-2-2b-it	299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8	gemma (base repo gated; GGUF used is a public repackaging)	#176
tiny_starcoder_py	starcoder	bigcode/tiny_starcoder_py	8547527bef0bc927268c1653cce6948c5c242dd1	BigCode OpenRAIL-M	#178
phi15	phi2	microsoft/phi-1_5	77aa61eeac94fbf33d492b9f2744c98b42d5b5eb	Microsoft Research License	#179
stablelm2_16b	stablelm	stabilityai/stablelm-2-1_6b	f499ead74c53749bd93cebc6ce8bc0d7bdf1eaef	StabilityAI non-commercial research	#179
internlm2_5_18b	internlm2	internlm/internlm2_5-1_8b-chat	57dcbc5455bf050d27a80f0ebd7808387caa714d	Apache-2.0	#181
starcoder2_3b	starcoder2	bigcode/starcoder2-3b	733247c55e3f73af49ce8e9c7949bf14af205928	BigCode OpenRAIL-M	#181
decilm7b	deci	Deci/DeciLM-7B-instruct	4adc7aa9efe61b47b0a98b2cc94527d9c45c3b4f	Apache-2.0	#182
minicpm5_1b	minicpm	openbmb/MiniCPM5-1B	4e9de7a0778dc1c362e983e6858f0e77542cbdca	Apache-2.0	#182
falcon7b	falcon	tiiuae/falcon-7b	ec89142b67d748a1865ea4451372db8313ada0d8	Apache-2.0	#183
afm45b	arcee	arcee-ai/AFM-4.5B	e84d19711eca5817c1608cb344d931bf05c966dd	Apache-2.0	#183
codeshell_7b	codeshell	WisdomShell/CodeShell-7B-Chat	3cb06f589b7b1e2f8e728c77280b1114191d24de	CodeShell License	#185
cohere2_7b	cohere2	CohereLabs/c4ai-command-r7b-12-2024	4f3d0aa6856e322f2f4480fe65420d5d53d297b8	CC-BY-NC-4.0 (repo gated)	#185
ernie45_03b	ernie4_5	baidu/ERNIE-4.5-0.3B-PT	b565cf6caebdb7a1eadf00100857b1ed5e044f12	Apache-2.0	#185
plamo3_nict_8b	plamo3	pfnet/plamo-3-nict-8b-base	7adeb6e9fc79016d4e99e2e3cb8fde6441b42366	PLaMo Community (repo gated, non-commercial)	#186
refact16b	refact	smallcloudai/Refact-1_6B-fim	a2b605b6d76ce09732ef6f54265b3c0a8393c08e	BigScience OpenRAIL-M	#186
maincoder_1b	maincoder	Maincode/Maincoder-1B	088ec98640bdeb105f46a9ef6a1370ed5d0d2ea5	Apache-2.0	#187
xverse7b	xverse	xverse/XVERSE-7B-Chat	60acc8c453c067b54df88be98bfdf60585ab5441	Apache-2.0	#188
bloom560m	bloom	bigscience/bloom-560m	ac2ae5fab2ce3f9f40dc79b5ca9f637430d24971	BigScience RAIL-1.0	#189
pythia410m	gptneox	EleutherAI/pythia-410m	9879c9b5f8bea9051dcb0e68dff21493d67e9d4f	Apache-2.0	#189

One more, mpt7b (arch mpt, ALiBi), is real and recipe'd on #187 but
deliberately has no claim file yet -- mosaicml/mpt-7b returns 401 on
both its API and its plain webpage for anonymous access, so we can't get a
real pinned revision to put in the claim. Will follow up once that's
resolved rather than ship an empty/floating revision.

Two models we already built and recipe'd are intentionally not in this
list: internlm3_8b and baichuan2_7b both resolve to plain llama arch
(confirmed live against the real GGUF loader), so per the track rules they
wouldn't create additional credit over the seed llama identity -- no
claim filed for either, on purpose.

Every claim above is a pure addition (new root + new claim file only,
nothing outside those paths, no artifacts.json/benchmark_config.json
edits in the claim itself) so none of this should trip anything you're
flagging as wrongly inserted. Whenever you get to reviewing identities, all
20 are ready to validate as soon as their identity_id lands in
data/model-port-identities.json.

[DarthCeltic]

Draft validation contracts for all 20 candidates above (structurally complete, real provenance data, PPL/correctness fields honestly marked pending a real host/board run rather than fabricated): #202

[DarthCeltic]

One more, found on a re-audit -- distilbert_sst2 (identity bert, execution family LLM_ARCH_BERT): distilbert-base-uncased-finetuned-sst-2-english @ 714eb0fa89d2f80546fda750413ed43d93601a13, Apache-2.0. First non-causal-LM (encoder classification) port on the board. Claim now on PR #172. (roberta_sst2, same PR, shares the identical arch -- not filed as a second claim, per the one-credit-per-family rule.)