Upload tpy07-v1_base model

checkpoints/attr_aliases.yaml

@@ -0,0 +1,72 @@
+# Fixed 2-letter aliases for common Python attribute names.
+#
+# Any `Attribute.attr` node whose name appears here is emitted as the corresponding 2-letter
+# alias. Attrs not in this table fall back to being emitted as a raw single-letter stream,
+# which often pushes the sample past the 128-token cap and gets it filtered out.
+#
+# Invariants enforced by the Vocab loader:
+#   1. No alias equals a 2-letter Python keyword (`as`, `if`, `in`, `is`, `or`).
+#   2. No alias appears in both this file and `builtin_aliases.yaml`.
+#
+# Aliases are picked as distinctive mnemonics; collisions resolved by picking distinct
+# second letters within a functional group.
+
+# --- list / sequence methods ---
+append: aa
+extend: xt
+insert: ti  # "in" is a keyword, use ti (ins-erT + I)
+remove: mv
+clear: cw
+count: ct
+
+# --- dict / set methods ---
+update: ud
+setdefault: sf
+pop: pp
+popitem: pi
+items: jm  # "it" would collide with iter's alias; use j as mnemonic for pl-j-urals
+keys: kk
+get: gu
+values: vu
+
+# --- sorting / reversing (distinct from builtins sorted/reversed) ---
+sort: so
+reverse: vv  # "rv" is taken by reversed()
+
+# --- string methods ---
+find: fi
+split: pl
+join: jn
+strip: tq
+lower: lw
+upper: ur
+replace: rl
+startswith: sw
+endswith: ew
+encode: nc
+decode: dk  # "de" feels too generic
+format: ft
+
+# --- file / io methods ---
+read: rd
+write: wt
+readline: rq
+readlines: rw
+close: cg
+seek: sq
+tell: tu
+flush: fu
+
+# --- object attributes ---
+name: nm
+value: vl
+data: da
+size: sz
+length: lh
+shape: sk
+dtype: dy
+
+# --- common numeric attrs (ndarray / tensor) ---
+mean: mj
+std: sj
+var: va

checkpoints/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0a7e592596866548a61b317bdeebff52fd11d003ba96d02c02ea7f9e610beab
+size 734545

checkpoints/builtin_aliases.yaml

@@ -0,0 +1,80 @@
+# Fixed 2-letter aliases for common Python built-in names.
+#
+# Each alias costs 2 alphabet tokens per use but preserves the full 26-slot user-identifier
+# table. The Vocab loader enforces two invariants:
+#
+#   1. No alias equals a 2-letter Python keyword (`as`, `if`, `in`, `is`, `or`) — those
+#      would be retokenized as keywords during decoding.
+#   2. No alias appears in both this file and `attr_aliases.yaml` — the two tables share a
+#      single inverse-lookup namespace at decode time.
+#
+# Grow this list when your dataset uses built-ins not yet covered. Aliases are chosen as a
+# distinctive 2-letter mnemonic of the name.
+
+# --- common functions ---
+print: pr
+len: le
+range: ra
+int: nt
+str: st
+list: ls
+dict: dc
+set: et
+tuple: tp
+type: ty
+map: mp
+filter: fl
+zip: zp
+enumerate: em
+sorted: sr
+reversed: rv
+sum: sm
+min: mn
+max: mx
+abs: ab
+any: ay
+all: aw
+open: op
+repr: rp
+hash: hh
+bool: bl
+float: fo
+bytes: by
+input: pu
+round: rn
+divmod: dm
+pow: pw
+iter: tr
+next: nx
+object: ob
+hex: hx
+oct: oc
+bin: bn
+chr: cr
+ord: od
+format: ff
+callable: cb
+
+# --- common methods masquerading as built-ins ---
+isinstance: ii
+hasattr: hs
+getattr: ga
+setattr: sa
+super: su
+property: pt
+staticmethod: sc
+classmethod: cm
+
+# --- exceptions ---
+ValueError: ve
+TypeError: te
+IndexError: ix
+KeyError: ke
+StopIteration: si
+Exception: xe
+NotImplementedError: ni
+RuntimeError: re
+AttributeError: ae
+ZeroDivisionError: zd
+FileNotFoundError: fx
+OSError: oo

checkpoints/manifest.json

@@ -0,0 +1,6 @@
+{
+  "fingerprint": "e2a95ef79baa37962dfec3897f03b8b54774e8c6e923b8c7badd1d77ca326ff1",
+  "epoch": 20,
+  "val_loss": 0.1479419188634042,
+  "encoder_name": "jhu-clsp/ettin-encoder-17m"
+}

checkpoints/vocab_98.yaml

@@ -0,0 +1,141 @@
+# Default 98-token vocabulary.
+#
+# Token IDs are assigned by concatenating the ordered category lists in the order they appear
+# below. IDs are therefore content-addressable and stable across reloads. The Vocab loader
+# computes a sha256 "fingerprint" over the canonical YAML, which is embedded in every checkpoint
+# to detect silent mismatches at inference time.
+#
+# Categories (ordered):
+#   specials   (3)   BOS, EOS, PAD
+#   structural (4)   NEWLINE, INDENT, DEDENT, STR_QUOTE
+#   keywords   (35)  Python reserved words
+#   alphabet   (26)  single-letter identifiers a..z
+#   digits     (10)  0..9
+#   operators  (23)  punctuation and multi-char ops (<=, >=, ==, !=, **)
+#
+# Total: 3 + 4 + 35 + 26 + 10 + 23 = 101.
+#
+# NOTE: The design doc states 98 tokens. The specials are counted separately in the spec
+# (BOS/EOS/PAD = 3 specials, not part of the "98"). When we say "98-token vocabulary" we mean
+# the 98 content-bearing tokens (35 keywords + 26 alphabet + 10 digits + 23 operators +
+# 4 structural). Total model vocabulary size including specials is 101; this is what
+# `Vocab.size` reports, since specials have real IDs that the decoder emits.
+
+version: 1
+
+categories:
+  specials:
+    - PAD
+    - BOS
+    - EOS
+  structural:
+    - NEWLINE
+    - INDENT
+    - DEDENT
+    - STR_QUOTE
+  keywords:
+    - "False"
+    - "None"
+    - "True"
+    - and
+    - as
+    - assert
+    - async
+    - await
+    - break
+    - class
+    - continue
+    - def
+    - del
+    - elif
+    - else
+    - except
+    - finally
+    - for
+    - from
+    - global
+    - if
+    - import
+    - in
+    - is
+    - lambda
+    - nonlocal
+    - not
+    - or
+    - pass
+    - raise
+    - return
+    - try
+    - while
+    - with
+    - yield
+  alphabet:
+    - a
+    - b
+    - c
+    - d
+    - e
+    - f
+    - g
+    - h
+    - i
+    - j
+    - k
+    - l
+    - m
+    - n
+    - o
+    - p
+    - q
+    - r
+    - s
+    - t
+    - u
+    - v
+    - w
+    - x
+    - y
+    - z
+  digits:
+    - "0"
+    - "1"
+    - "2"
+    - "3"
+    - "4"
+    - "5"
+    - "6"
+    - "7"
+    - "8"
+    - "9"
+  operators:
+    - "("
+    - ")"
+    - "["
+    - "]"
+    - "{"
+    - "}"
+    - ":"
+    - ","
+    - "."
+    - "="
+    - "+"
+    - "-"
+    - "*"
+    - "/"
+    - "%"
+    - "<"
+    - ">"
+    - "!"
+    - "=="
+    - "!="
+    - "<="
+    - ">="
+    - "**"
+
+max_user_identifiers: 26
+max_string_length: 20
+
+string_charmap:
+  space: s
+  underscore: u
+  fallback: "?"

config.yaml

@@ -0,0 +1,73 @@
+encoder:
+  name: jhu-clsp/ettin-encoder-17m
+  max_length: 512
+  dtype: bfloat16
+  trust_remote_code: false
+decoder:
+  d_model: 64
+  n_layer: 2
+  n_head: 4
+  d_ff: 256
+  max_seq_len: 128
+  prefix_len: 4
+  dropout: 0.0
+  tie_embeddings: true
+  cross_attention: false
+  enc_memory_len: 128
+  pos_encoding: learned
+  mlp_type: gelu
+  copy_head: false
+data:
+  hf_name: parquet
+  data_files: /tmp/tpy_build/07_deduped.parquet
+  split: train
+  val_fraction: 0.02
+  max_identifiers: 26
+  max_tokens: 128
+  instruction_field: instruction
+  code_field: response
+  stdlib_whitelist_path: null
+  num_proc: 8
+  cache_dir: .cache/tiny_decoder/tpy07_v1
+  drop_stub_bodies: false
+  pack: false
+  pack_len: 512
+vocab:
+  path: configs/vocab_98.yaml
+  builtin_aliases_path: configs/builtin_aliases.yaml
+  attr_aliases_path: configs/attr_aliases.yaml
+train:
+  batch_size: 128
+  epochs: 50
+  lr: 0.0003
+  warmup_frac: 0.05
+  weight_decay: 0.1
+  grad_clip: 1.0
+  early_stop_patience: 5
+  seed: 0
+  log_every: 50
+  eval_every_epochs: 1
+  wandb_project: null
+  tensorboard_dir: null
+  weighted_sampling: false
+  init_from: null
+  pretrain_sampling: short
+  amp_dtype: bf16
+  compile: false
+  bench_every_epochs: 0
+  bench_benchmark: both
+  bench_max_problems: null
+  bench_num_samples: 5
+  bench_temperature: 0.8
+  bench_pass_k: 1,5
+  bench_timeout: 3.0
+  bench_num_workers: 16
+  bench_max_new_tokens: null
+eval:
+  split: val
+  max_samples: null
+  compute_exact_match: true
+  compute_syntactic_validity: true
+run:
+  dir: runs/tpy07/v1_base
+  save_every_epochs: 1