Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

Enhance finetuning process with progress logging and update HTML hints for clarity

HenryNdubuaku2ec297247e56
diff --git a/needle/model/finetune.py b/needle/model/finetune.py
index 75e20b6..7a3a7fe 100644
--- a/needle/model/finetune.py
+++ b/needle/model/finetune.py
@@ -257,13 +257,18 @@ def merge_lora(params, lora, scale):
return unflatten_dict(flat)
-def finetune_local(args):
+def finetune_local(args, progress=None):
import jax
import jax.numpy as jnp
import optax
from .run import load_checkpoint
from .architecture import SimpleAttentionNetwork
+ def emit(msg):
+ print(msg, flush=True)
+ if progress:
+ progress(msg)
+
base_path = args.checkpoint or DEFAULT_BASE
data_path = args.jsonl_path
if getattr(args, "generate", 0):
@@ -276,13 +281,13 @@ def finetune_local(args):
seqs, masks = load_jsonl(data_path, tokenizer, args.max_len)
if len(seqs) == 0:
raise SystemExit("no usable examples in " + data_path)
- print(f"training on {len(seqs)} examples, seq_len {args.max_len}")
+ emit(f"training on {len(seqs)} examples, seq_len {args.max_len}")
model = SimpleAttentionNetwork(config)
paths = lora_target_paths(params)
scale = args.lora_alpha / args.lora_rank
lora = init_lora(params, paths, args.lora_rank, jax.random.PRNGKey(0))
- print(f"LoRA rank {args.lora_rank} on {len(paths)} weight groups")
+ emit(f"LoRA rank {args.lora_rank} on {len(paths)} weight groups (compiling...)")
optimizer = optax.adamw(args.lr)
opt_state = optimizer.init(lora)
@@ -300,6 +305,10 @@ def finetune_local(args):
return optax.apply_updates(lora, updates), opt_state, loss
batch, count = args.batch_size, len(seqs)
+ steps_per_epoch = -(-count // batch)
+ total_steps = args.epochs * steps_per_epoch
+ every = max(1, total_steps // 50)
+ step_i = 0
for epoch in range(args.epochs):
order = np.random.permutation(count)
last = 0.0
@@ -308,7 +317,10 @@ def finetune_local(args):
lora, opt_state, loss = train_step(lora, opt_state,
jnp.asarray(seqs[idx]), jnp.asarray(masks[idx]))
last = float(loss)
- print(f"epoch {epoch + 1}/{args.epochs} loss {last:.4f}")
+ step_i += 1
+ if step_i % every == 0:
+ emit(f"epoch {epoch + 1}/{args.epochs} step {step_i}/{total_steps} loss {last:.4f}")
+ emit(f"epoch {epoch + 1}/{args.epochs} loss {last:.4f}")
os.makedirs(args.checkpoint_dir, exist_ok=True)
out = args.out or os.path.join(args.checkpoint_dir, "needle_lora.pkl")
diff --git a/needle/playground/index.html b/needle/playground/index.html
index cc2c02c..07dc83f 100644
--- a/needle/playground/index.html
+++ b/needle/playground/index.html
@@ -62,7 +62,7 @@
<button class="btn-primary" id="sendBtn" onclick="send()">Run</button>
<button class="btn-ghost" id="resetBtn" onclick="newChat()" title="Start a fresh conversation">New chat</button>
</div>
- <p class="hint">Runs continue the same conversation, so you can ask follow-ups. Picking a preset or "New chat" resets the session.</p>
+ <p class="hint">Keep asking follow-up questions against the same tools - each Run is a fresh turn. Pick a preset or "New chat" to switch tools.</p>
</div>
<div class="result-section">
<label class="panel-label">Conversation</label>
diff --git a/needle/playground/server.py b/needle/playground/server.py
index 9e87be1..682a9d1 100644
--- a/needle/playground/server.py
+++ b/needle/playground/server.py
@@ -28,11 +28,13 @@ class Engine:
self.tools_json = "[]"
def complete(self, tools_json, query):
- from .. import Needle
+ from .. import Needle, _lib
with self.lock:
if self.agent is None or tools_json != self.tools_json:
self.agent = Needle(tools=tools_json, weights=self.weights)
self.tools_json = tools_json
+ else:
+ _lib().needle_reset()
return self.agent.complete(query)
def reset(self):
@@ -55,6 +57,12 @@ class Engine:
_FT = {"running": False, "step": "", "log": [], "checkpoint": None, "error": None}
+def _log(msg):
+ _FT["log"].append(msg)
+ if len(_FT["log"]) > 100:
+ del _FT["log"][:-100]
+
+
def _finetune_worker(tools_json, api_key, samples, engine):
import types
try:
@@ -64,7 +72,7 @@ def _finetune_worker(tools_json, api_key, samples, engine):
tools = json.loads(tools_json)
rows = generate_dataset(
tools, samples, api_key=api_key,
- progress=lambda done, total: _FT["log"].append(f"generated {done}/{total}"))
+ progress=lambda done, total: _log(f"generated {done}/{total}"))
data_path = str(_DOWNLOADS / "needle_playground_data.jsonl")
with open(data_path, "w") as handle:
for row in rows:
@@ -75,7 +83,7 @@ def _finetune_worker(tools_json, api_key, samples, engine):
finetune_local(types.SimpleNamespace(
jsonl_path=data_path, checkpoint=None, epochs=3, batch_size=16, lr=1e-4,
lora_rank=16, lora_alpha=32.0, max_len=1024, generate=0, model=None,
- checkpoint_dir=str(_DOWNLOADS), out=adapter))
+ checkpoint_dir=str(_DOWNLOADS), out=adapter), progress=_log)
_FT["step"] = "building"
out = str(_DOWNLOADS / "needle_tuned.cact")
diff --git a/tests/test_finetune.py b/tests/test_finetune.py
index 71c52e9..20067ed 100644
--- a/tests/test_finetune.py
+++ b/tests/test_finetune.py
@@ -37,8 +37,11 @@ def test_finetune_writes_adapter(tiny_checkpoint, tmp_path):
data = tmp_path / "data.jsonl"
_write_data(data)
out = tmp_path / "adapter.pkl"
- finetune_local(_finetune_args(data, tiny_checkpoint, out, tmp_path / "ck"))
+ progress = []
+ finetune_local(_finetune_args(data, tiny_checkpoint, out, tmp_path / "ck"),
+ progress=progress.append)
+ assert any("loss" in m for m in progress)
assert out.exists()
with open(out, "rb") as handle:
adapter = pickle.load(handle)